Skydio Targets Zero-Pilot Fleets With Asymmetric F10 and MegaDock

Skydio推出非对称F10与MegaDock，发力零飞手无人机机队 只有借助边缘感知与自动化机巢将人机飞手比例降趋近于零，自主空中机器人才能扩展为全天候公共基础设施。 Juraj Kabzan介绍了Skydio从消费级无人机迈向全天候运行的全自主企业级机队机器人的技术演进。通过网页浏览器现场演示远程多无人机协同编排，他讲解了在城市环境中处理实时避障、视觉惯性里程计和预测性追踪的边缘与云端双重自主架构。本次研讨会重点介绍了无人机作为第一响应者（DFR）在美国主要警察部门的实地部署、非对称F10 Lightrunner及机械臂MegaDock背后的工程设计，以及该公司旨在实现自然语言机队控制的视觉语言基础模型研发路线图。 Skydio已从消费级摄像设备演变为关键企业基础设施，向4,000家客户部署了超过60,000架无人机，超过2800万美国人生活在现役第一响应者机巢方圆两英里范围内。 (6:35) 为解决城市峡谷中的视线遮挡问题，Skydio采用两阶段强化学习策略来预判目标运动，并主动绕过建筑物转角以维持持续追踪。 (24:30) 非对称F10 Lightrunner与机械臂MegaDock相结合，实现了100英里/小时的冲刺速度和60分钟的续航时间，同时通过机械臂抓取回收消除了起落架带来的额外寄生重量。 (40:01) 安全攸关的感知与飞行控制在本地的NVIDIA Jetson Orin NX上执行，而数十亿参数的开放词表视觉语言模型则通过冗余的5G及无线电链路在云端运行。 (33:51)

Juraj Kabzan argues that aerial robotics are evolving from consumer gadgets into 24/7 critical civil infrastructure, driven by complete vertical hardware integration, automated docking, and an autonomy architecture designed to shift human pilots from manual control to high-level fleet orchestration. Scaling Public Safety and Fleet Orchestration (18:26) Drone as a First Responder (DFR) operations have demonstrated concrete tactical value across municipal agencies, with telemetry showing drones arrive on scene first 52% of the time. Incidents cited include the Bernalillo County Sheriff's Office safely de-escalating an armed encounter after aerial feeds revealed children firing a stolen weapon, and Oklahoma City police rapidly locating an injured pedestrian along rugged railroad tracks. Kabzan notes that over 28 million Americans now reside within two miles of a Skydio docking station. However, scaling municipal operations is constrained by staffing bottlenecks under traditional 1:1 pilot-to-drone ratios. A simulation of Clark County's 500,000 emergency calls indicated that full geographic coverage requires 160 docks, which would necessitate 160 dedicated human pilots. To address this limitation, Skydio introduced multi-vehicle supervision allowing a single pilot to manage four drones concurrently, with an engineering objective to push human supervision ratios near zero by shifting pilots into high-level orchestrators issuing abstract mission tasks. Partitioned Edge-Cloud Autonomy and Sensing (33:51) Autonomous flight in complex urban environments requires balancing strict Size, Weight, and Power (SWaP) constraints on the airframe against high computational demands. Skydio partitions its architecture between edge processors and cloud infrastructure. Onboard NVIDIA Jetson Orin NX units execute high-frequency, safety-critical loops—including 360-degree vision-based obstacle avoidance and visual-inertial odometry—ensuring fail-safe execution even if communication links drop. Heavyweight vision-language foundation models run in the cloud, processing live video feeds streamed via Connect Fusion+, a network system that aggregates multiple commercial cellular carriers and local radio channels. To navigate dense downtown corridors without positioning failures, the platform resolves urban GNSS multipath distortion through camera-aided sensor fusion. Because standard GPS signals bounce off high-rise structures, the autonomy stack projects expected satellite line-of-sight vectors into calibrated camera frames, identifying and systematically culling signals that correspond to visually obstructed satellites before they corrupt global pose estimation. Aerodynamic Innovation and Predictive Tracking (40:01) Conventional fixed-wing drones offer superior transit speed and loiter endurance over multirotors but historically suffered from two flaws: requirement of large ground crews for recovery, and camera occlusion caused by the inboard wing dipping into the sensor's line of sight during banked turns. To resolve this trade-off, Skydio developed the F10 Lightrunner, an asymmetric tailsitter reaching dash speeds of roughly 100 mph with 60 minutes of endurance. The airframe mounts its sensor payload on the far left wingtip and uses a trajectory planner constrained to left-hand loitering orbits. Ground support is automated through the MegaDock, which uses a robotic manipulation arm to launch and recapture the vehicle. For tracking moving subjects on the ground, Skydio deployed "Shadow," a two-stage reinforcement learning policy. Rather than passively following target telemetry, the policy recognizes deceleration cues, accelerates forward, and positions the aircraft overhead prior to street turns to preserve an uninterrupted visual line of sight around tall buildings. Environmental Boundaries and Operational Transparency (56:27) Despite automated navigation, physical flight physics and perception boundaries impose strict operational limits. While Skydio airframes are waterproof, active obstacle avoidance must be disabled during rain because falling droplets create transient sensory noise that local voxel mapping algorithms misinterpret as physical obstacles, causing erratic flight maneuvers. In GPS-denied or actively jammed defense environments, platforms like the tactical X10D substitute GNSS with altitude-optimized visual odometry, which inevitably accumulates positional drift over distance but remains sufficiently accurate to guide aircraft back to their launch points. Regulatory compliance and societal integration represent additional operational hurdles. Deploying multi-drone operations Beyond Visual Line of Sight (BVLOS) requires tailored FAA Part 107 regulatory waivers. To build public trust, Skydio established a Transparency Dashboard allowing participating public safety agencies to publish audit logs and flight paths for public scrutiny. While Skydio aims to eliminate manual piloting in favor of natural language mission orchestration, current autonomy remains bounded: human operators must intervene whenever tracking targets are fully obstructed, and adverse weather or sensor limitations enforce operational flight ceilings. Realizing scalable, pilotless operations depends on ongoing work in model distillation to onboard edge compute and expanding regulatory BVLOS frameworks.

Skydio自主系统架构师Juraj Kabzan在斯坦福研讨会上指出，无人机已从消费级摄影工具跃升为全天候运转的关键民生与安全基础设施。依托美国本土制造的全栈垂直整合、端云分工架构及数据飞轮，Skydio正推动机队操作从“一人一机”向以高层意图驱动的“趋近零人工编排”演进。 垂直整合与机队自主编排体系 (6:35) Kabzan强调，Skydio的核心壁垒在于全栈垂直整合：自主设计停机坪（Dock）机械结构、注塑机身并于美国本土制造，自研低层嵌入式电机控制、端侧视觉AI感知以及云端机队指挥系统（Remote Flight Deck）。目前公司拥有约1000名员工，已向4000多家企业及公共安全客户交付超6万架无人机。在警务与巡检场景中，传统“一人一机”操作模式构成了扩张瓶颈；例如对克拉克县约50万起应急警情进行全域覆盖需布局160个机巢，配备等量飞官成本极高。为此，Skydio推行一人监管多机的编排架构，目前一名飞官可同时监管4架以上无人机，最终目标是将飞官角色精简为仅下达任务意图的高层编排者，依靠机载与云端协同实现航线生成及近距空域自动解冲。 第一响应者与基础设施实战效益 (12:00) 无人机作为第一响应者（DFR）已展现出量化治理价值。数据显示，全美已有超2800万人生活在Skydio机巢2英里半径内；丹佛与拉斯维加斯警务实战表明，无人机在52%的出警事件中先于地面巡逻车抵达现场。旧金山部署DFR后总体犯罪率降低30%，纽约市警局则借此干预了200余起地铁车顶穿行的高危险情。在具体案例中，伯纳利欧县警方通过无人机俯瞰画面及时发现持枪涉案人员实为幼童，避免了致命武力使用；俄克拉荷马城警方则依靠沿铁路巡查的无人机迅速定位重伤行人挽救生命。在公共事业领域，永久机巢被部署于变电站周边，操作员通过浏览器即可启动巡查，曾成功排查出电力配电杆塔的热灼伤缺陷，避免了大面积电网故障。 端云协同计算与复杂环境鲁棒感知 (23:45) 针对城市峡谷环境中的定位失效难题，Skydio开发了传感器融合剔除机制：系统掌握全天卫星星历向量，并将其投影至机载标定相机视场中，一旦视觉识别到建筑遮挡，便在滤波算法中主动剔除该卫星的多径反射信号。机体搭载的NVIDIA Jetson Orin NX芯片负责处理360度视觉避障、视觉惯性里程计（VIO）等高频安全控制回路，而超大参数量的开放词表模型与复杂语义理解则部署于云端。在跟踪逃逸车辆的“Shadow”功能中，系统采用端到端强化学习策略：当被跟踪车辆减速并准备转弯时，算法会预判轨迹提前加速并掠至车辆上方，以规避后续因高层建筑遮挡导致的视线丢失。仿真策略直接基于数字表面模型（DSM）训练，由于真实机载系统采用同构地图表征，大幅收窄了虚实迁移鸿沟。 异构机型与气动工程突破 (40:01) 为兼顾高速到达与长时间留空，Skydio打破常规对称机翼设计，开发了非对称尾座式飞机F10 Lightrunner。传统固定翼飞机在低空盘旋时机翼会遮挡机腹中置云台视线，F10则将相机云台移至外侧机翼末端，配合统一的模型预测控制（MPC）轨迹规划器强制执行左侧盘旋，确保机身永远不阻挡镜头。该机型最高冲刺时速约100英里，续航达60分钟，并在20英里/小时的低速临界点自动转入垂直悬停。配套的MegaDock利用高精度机械臂实现垂直起降与回收收纳，使飞行器免除了传统垂直起降机构与起落架的死重。针对管道与室内等GPS拒止狭窄空间，配备全包覆防撞护罩的双鱼眼相机机型R10，结合NightSense红外补光套件，可实现全暗环境下的纯视觉自主穿行。 数据飞轮、适航合规与未来演进 (45:11) Kabzan指出，Skydio拥有全球规模最大的飞行机器人真实运行数据集，在客户脱敏授权下，遥测与传感器数据持续回流以驱动模型迭代。在合规层面，公司协助客户申请FAA Part 107框架下的超视距（BVLOS）飞行豁免，并上线透明度仪表盘（Transparency Dashboard）公开警务飞行轨迹以消除公众隐私疑虑。针对恶劣气象，机巢虽内嵌气象监测，但在雨天必须主动关闭机载主动避障功能，因下落雨滴会被局部点云算法误构为障碍物而诱发异常机动；大风仍是核心物理限制。未来，Skydio计划通过多通道通信聚合技术（Connect Fusion+）保障视频回传，并探索将视觉-语言-动作（VLA）模型蒸馏压缩至端侧，实现仅通过自然语言意图直接下达端到端自主飞行任务。 Skydio展示了从硬件自研、自主避障到云端编排的完整自主系统闭环，但其全天候运作能力仍受限于极端气象与雨天传感器噪点等物理法则。随着多机编排比例向趋近零人工推进，如何在复杂遮挡导致的目标彻底丢失、网络瞬断以及高动态城市空域交汇中厘清人机接管权责，仍是该体系规模化落地必须应对的核心挑战。

[20:22] Juraj Kabzan: Maps are never up to date.

[0:13] Uh, thank you for having me. Thank you [0:16] for allowing Skydio to present. And [0:18] yeah, I'll I'll tell you more about what [0:19] we have been up to for the last few [0:21] months and months and years. [0:23] And but before I actually go into the [0:26] you know the presentation stuff and [0:27] slide, I I do like to start with just [0:29] like giving you a glimpse of what [0:30] actually we have built. Um, so this is a [0:33] something that we call uh [0:35] um [0:35] a remote flight deck. So it's our main [0:37] user interface that you can log into or [0:39] anyone in the world you can log into and [0:41] you see sort of all of our all of these [0:42] systems across your across your [0:45] organization. In this case, you know, I [0:47] have here a bunch of docking station, [0:48] bunch of drones. I will click on on one [0:51] particularly that is you know dock 33 [0:53] that is in the that is in our in our [0:56] headquarters down in San Mateo, very [0:58] close from here. What you can see on the [0:59] map is sort of areas, you know, where we [1:01] can and cannot fly. Obviously, you know, [1:03] the red areas represent a like a um [1:06] maximum altitude and you cannot fly [1:07] where you cannot fly, mainly because of [1:09] the San Francisco San Francisco airport [1:12] or other ones that perhaps have limited [1:14] uh limited altitude. But uh without sort [1:16] of uh talking further, let me just enter [1:19] enter cockpit in one of these uh one of [1:22] these docking stations. Uh again, this [1:24] is completely live. This is not a video. [1:26] I'm doing everything in in in a live in [1:27] in live. So let's hope everything works [1:29] out. But yeah, these are sort of the [1:31] docking stations that are on the on our [1:33] on the roof of our headquarters and I'm [1:35] right now taking off with one of them. [1:38] And I will do my best piloting them from [1:41] a touchpad. It's a interesting [1:43] experience, but I'll I'll I'll do my [1:44] best. [1:45] Um hopefully our underlying autonomy [1:47] system will allow me to actually still [1:49] do it efficient efficiently. So yeah, I [1:51] am I am right now in there. As you can [1:53] see, I'm moving it left and right and I [1:55] can go uh change to something pointer [1:57] lock. or I'm doing is moving sort of my [1:59] my on the mouse pad and I'm [2:02] I'm sort of controlling the drone flying [2:04] around. I can see exactly on left side [2:06] of you know where I'm looking at it on [2:08] the map. [2:09] And [2:10] yeah, I can look around in our in our [2:12] headquarters see who is in the office [2:13] and who isn't. So you know, we can [2:16] we can have the conversation afterwards. [2:18] But let's say that you know, there is [2:20] something something interesting [2:21] happening with this car. So I'm going to [2:22] just just looking look at it. And you [2:25] know, I'm going to pick another drone [2:26] because you know, [2:27] there's always a scenario when you when [2:29] you when something critical is happening [2:31] that you want to keep view on something [2:33] while start to do something else. So I [2:35] looked I I enabled something called [2:36] share feature called shadow while while [2:39] the drone is sort of locked on that one. [2:41] On that particular car, I logged in to a [2:43] different drone number 44 and I can [2:48] I can take off with another drone. [2:51] While I will do that I can of course [2:54] still go back to the original drone and [2:56] control it in a manual way further, but [2:57] I'll wait once uh until it sort of takes [3:00] off. And as you can see here, there are [3:02] sort of you can I I have I have the map [3:04] area. These red or or orange orange [3:07] boxes represent sort of geo fences where [3:09] I cannot fly, but I can in this case [3:11] just click on the map and [3:14] and commandeer the drone and just say [3:15] fly here. And while it is flying, I [3:17] technically don't even have to [3:18] necessarily look. It's going to [3:19] completely completely do everything [3:21] autonomously. But I can while it flies [3:23] here, I can go back to the first drone. [3:26] Rotate around, see whether we have we we [3:28] are lucky enough to have a car. I see [3:30] car here. So let me just look at it [3:33] and that one is going to start tracking [3:34] that particular vehicle. I can go while [3:36] it tracks, I can go back to this one and [3:39] continue sort of doing [3:40] doing whatever I need whatever work I [3:42] need to do with the first drone. [3:44] And yeah, while while all of this [3:47] happening, I could if I want to I [3:49] actually let me try. I'll take another [3:51] drone. This is going to be number 88 [3:54] that I think is going to be completely [3:55] on the other side side of the [3:58] on the other side of the country. [4:00] I will stop the first one this tracking [4:02] the vehicle because we don't want to the [4:04] drone to fly with our geo fences so we [4:06] always make sure that the the drones can [4:08] fly only in a certain area and I know [4:10] that if the car goes further it will it [4:11] will go out of them so I'll just stop [4:13] it. [4:14] But [4:15] I'll go I'll now move to completely [4:17] different part of the country. [4:18] This is [4:20] this one is in the Colorado so that's [4:23] our another testing site which we have. [4:26] That I can again just I I change to this [4:28] drone and I can fly this drone [4:30] completely seamlessly. [4:32] Fly around change the altitude and again [4:34] do some do some useful useful work with [4:37] this drone. [4:38] And now I can go back to again see all [4:41] these drones at the same time and if I [4:43] want I can you know if I'm done with [4:45] some some work that I've been doing I [4:47] can just send all the drones [4:49] to return return landing dock and I [4:52] don't have to pay it I don't have to [4:53] really control them anymore all of this [4:55] is happening completely autonomously. As [4:57] you can imagine in this particular [4:59] scenario we have two drones that are [5:00] returning you see the our headquarter [5:02] our roof on headquarters there is [5:04] there's bunch of these docking stations [5:06] already and you might have noticed one [5:08] of them actually got paused that's [5:10] something that we call our airspace [5:11] deconfliction so even though there are [5:13] multiple drones flying at the same time [5:15] in the same environment we auto [5:17] automatically pause the drone that is [5:19] sort of close to it such that they never [5:21] they never come to a collision. So it's [5:22] again completely autonomous you you [5:24] don't have to you don't have to pay [5:25] attention to it they deconflict itself [5:27] once the once the first drone let's say [5:29] land the second one is going to get [5:31] unblocked and [5:33] is going to land afterwards. [5:34] After sort of I am done with my mission [5:36] I get a post flight report that I can [5:38] sort of I can sort of share any insights [5:40] about the flight what I what useful work [5:43] I have done later. [5:44] And [5:46] that's about it so this this is sort of [5:48] again what you what we have just seen is [5:50] like three drones flying two of them in [5:52] our headquarters Samatel, one in one in [5:54] a Colorado flying just from here from [5:57] this laptop in completely autonomously. [5:59] And I had full faith in our underlying [6:01] autonomy system that even though if I [6:03] went in into a building, you know, it [6:05] would still it would not allow me to our [6:07] on board collision avoidance would [6:08] prevent me prevent me to collide with [6:11] anything. [6:12] So, that's a really a high-level feature [6:14] capability what you can do all of what [6:16] you can do with these drones. [6:18] And I'll jump back to [6:21] back to the presentation [6:23] to tell you more about, you know, what's [6:24] what what's what what we are seeing [6:27] where the most use of these drones are [6:28] on these all systems as well as touching [6:30] on the other, you know, other other [6:32] robotic platforms that we have been [6:33] working on. [6:35] So, yeah, [6:36] how many of you have heard of Skydio? [6:38] I'm curious. [6:39] Oh, nice. Awesome. That's great. I I [6:42] suspect many of you have heard of us [6:43] heard about us mainly from our consumer [6:45] days. So, this is sort of where we [6:46] started of course more than 10 years ago [6:49] with Skydio R1 and Skydio Skydio 2. [6:51] Those are the first two drones mainly [6:53] used for uh, [6:55] so taking pictures, taking very very [6:57] nice videos of of of people doing [6:59] doing doing sports. We were able to sort [7:01] of track very closely while avoiding [7:04] avoiding obstacles and keeping them [7:05] still line of sight. So, that's really [7:07] the foundation that we have been we have [7:09] been building on on on on for a long [7:11] time. [7:12] What we are doing uh, is is taking sort [7:13] of off the shelf sort of [7:15] commodity mobile CPUs as well as imaging [7:17] sensors. [7:19] Of course, making them custom more more [7:20] efficient and putting in on four motors [7:23] and propellers, putting the cutting edge [7:25] AI technology and building flying [7:26] robots. And that's really the the the [7:28] fun part working at Skydio. [7:30] A little bit of, you know, about the [7:31] company where we are right now. Founded [7:33] in 2014, we have more than 60,000 at [7:35] this point more drones shipped and more [7:38] at this point more than 4,000 4,000 [7:40] customers. So, it's really a [7:41] fairly matured. Yes. [7:44] Right around 1,000. [7:45] Around 1,000 around 1,000 employees. [7:47] Yeah. [7:48] The autonomy team itself is around 50 60 [7:50] people. [7:52] Um yeah, so the the main thing and I and [7:55] I think we have seen this across many [7:56] industries. The very important part and [7:58] we we see it as well is that at Skydio [8:00] we build everything from the dock [8:02] hardware to the drone hardware as well [8:04] as we do all the manufacturing in US by [8:06] by us as well as we build all the [8:08] low-level vehicle software, the autonomy [8:10] software, and the cloud software. And [8:12] really the fact that the the capability [8:14] or [8:14] the the fact that we built all of it [8:16] allows us to really optimize the whole [8:19] stack and and get the most of the last [8:21] juice of the last percentage of of a [8:23] performance out of these systems. And [8:25] that turns out to be incredibly [8:27] important, especially as you know, the [8:29] the usage of these system grows and the [8:31] reliability of the system that is [8:33] required to actually deliver useful work [8:35] goes up very very quickly. [8:37] This is actually a um an example of one [8:39] of our deployments. This is not [8:40] AI-generated video uh picture even [8:42] though it may look like it. This is uh [8:44] in Las Vegas in Las Vegas. These are [8:46] sort of the they have deployed bunch of [8:48] these hives as we call them. Hives is [8:50] sort of the one deployment in one [8:52] location of multiple docking stations. [8:54] And they have many of these across the [8:56] entire city. And then they have one area [8:58] that they call Arctic uh real-time crime [9:01] center that they have the pilots sitting [9:03] in and they can coordinate and also [9:05] uh teleoperate all of these all of these [9:07] drones from that one single location. [9:09] We are like these are already deployed [9:11] at scale and we are hearing there is [9:13] constant influx of success stories [9:15] across all our portfolios of customers. [9:17] This is just one example from Bernalillo [9:19] County Sheriff's Office where they heard [9:21] uh [9:22] where they got a call that there's [9:23] someone uh shooting a gun. [9:26] Then so they take it took a stand, they [9:27] deployed it, and when they actually [9:29] looked from different perspective, they [9:30] noticed these are actually kids who [9:31] stole a gun from from their parents. [9:34] Uh and this is a very very critical [9:37] information that was then relayed to [9:38] these police officers that were on the [9:39] on the site, and they managed to adjust [9:42] their response according to knowing this [9:44] information. They said themselves, and [9:46] once they knew it, then they disarmed [9:48] disarmed the kids, took the gun, and [9:50] de-escalated de-escalated the whole [9:51] situation. They said themselves that if [9:53] they didn't know this very critical [9:55] piece of information, however small it [9:57] may look like, the entire incident might [9:59] have might have turned in much worse [10:01] much worse direction. They might have [10:02] been killed or something something along [10:03] those lines. But, it's very important to [10:05] have the right information at the right [10:07] time, which really can save lives. [10:10] Similarly, from OKC PD, this was one of [10:13] the case that that they got a call that [10:16] that a person was hit by a train. They [10:18] immediately dispatched a drone, flew to [10:20] to the location where where they guess [10:22] where they guess it happened, but as you [10:24] can imagine, train tracks are very long. [10:26] So, they took a drone, they flew along [10:27] the train tracks, and very quickly they [10:29] were able to actually find that there [10:31] was indeed a person that who was hit by [10:33] a train, and they have identified [10:35] the person where it is, sent paramedics [10:37] there, and they were actually able to [10:39] save life of this person. Again, it's [10:40] very important that you get the right [10:42] information at the right time as quickly [10:43] as possible. [10:45] And lastly, well, perhaps you know, the [10:46] closest to us down from San Francisco, [10:49] they got a call of a stolen car. So, [10:50] again, dispatched one one drone. [10:52] Throughout this entire event, they [10:54] dispatched actually multiple drones to [10:55] cover to to keep an eye on this car for [10:57] entire period of time. They always had [10:59] it [11:00] had it in line of sight. This actually [11:02] incident that they said that they heard [11:03] of that is happening, but had never seen [11:05] or had it on the footage. In this case, [11:07] the the the [11:09] the person is stealing another license [11:10] plate from another car, sitting in [11:12] there, driving a little bit forward [11:15] further, and what he what what he does [11:16] then is actually exchange the plate. [11:18] Thankfully, he nicely shows it to the [11:20] drone, so we can very clearly see the [11:21] new license plate, you know, thank God. [11:24] Uh, so, then we always we we the drone [11:27] always kept this person in in the field [11:29] of view. Now, he's tinting his [11:32] his [11:33] uh windows. [11:34] And, you know, this is very critical [11:36] part that the police officials always [11:37] know where where the car is. And once [11:39] they choose that this is like a safe [11:41] time, they actually rolled out the [11:42] spikes uh and and got the car got the [11:45] car to stop and got the people to [11:47] custody. And then again, all everything [11:48] was de-escalated very fairly fairly [11:50] efficiently without anyone hurting and [11:53] any any sort of damage. So, that's it's [11:54] really incredible how the right right [11:57] information at the right time can change [11:58] the outcome. [12:00] And we are seeing statistics across all [12:01] our customers, you know, for example, in [12:03] San Francisco, they report 30% drop in [12:06] overall crime [12:07] overall crime since they deployed these [12:09] systems. [12:10] Uh New York, they have they have [12:12] intervened in more than 200 subway [12:14] serving incidents again uh saving lives. [12:17] And you know, these statistics keep [12:18] going. Denver Police Department, Las [12:20] Vegas, they are all all seeing that the [12:22] thing because the drone can can be [12:24] deployed so quickly from 911 call and [12:26] can fly to the destination where [12:28] actually this 911 call happens, they see [12:30] that 52% of the time drone is the first [12:33] thing that arrives to the scene. And as [12:35] you can imagine in the in the scenarios [12:37] that it matters the most, if you get a [12:39] if you get there earlier and you can [12:40] actually relay what's happening back to [12:42] the back to the coordinating officer, [12:45] uh this additional information may [12:46] adjust very sub- substantially what type [12:49] of response they will do. And this is [12:50] like extremely this is this turns out to [12:52] be extremely important [12:54] across all all our customers. [12:56] At this point of time, more than 28 [12:59] million Americans live within 2 miles of [13:01] a of a of a Skydio first responder. And [13:04] first uh uh first responder drone is [13:07] really is is a is a system that we call [13:09] uh that police officers are using these [13:11] docking stations to respond to 911 [13:12] calls. [13:14] Um the you know, on the other side [13:18] uh [13:21] Hopefully. [13:22] Cool. [13:24] On the other side, you know, I talked a [13:25] lot a lot about the police deployments, [13:27] this is just one one one vertical that [13:29] these systems are used for. This is a a [13:32] video, hopefully it loads, [13:34] from a a substation where So, this dog [13:36] is deployed right next to a substation [13:38] where people can where where where [13:41] utility uh [13:42] companies can operate them remotely, [13:44] take off and inspect the [13:46] inspect the the the substations for any [13:49] possible damage. And they can do it [13:51] proactively or reactively depending on [13:52] the situation, but it's very important [13:54] that they identify if there is any [13:56] possible any possible uh damage early on [13:59] such that they can perhaps fix it even [14:00] before some some damage occurs. In this [14:03] case, [14:04] in this case they have uh identified [14:06] actually a burn like a burned area on a [14:09] pole. And again, this information is [14:11] very important because if you can react [14:12] quickly, you can actually fix it before [14:14] any sort of further damage happens and [14:16] perhaps an area loses entirely power. [14:19] And this [14:20] um [14:22] And this information is again uh it's [14:23] important to get this information as [14:25] quickly as possible. [14:26] The [14:29] Uh [14:30] that's I think the same one. Perfect. [14:33] And a part of that, we are of course of [14:35] course supporting We have We have [14:36] shipped many many drones to the national [14:38] security customers [14:40] uh uh and our allies. [14:42] Um in terms of the you know, in terms of [14:44] the growth, we are seeing, you know, [14:45] especially since we have deployed this [14:47] X10 Plus docking station, our sales are [14:50] are are the the market is seeing that [14:52] the the systems can generate tremendous [14:54] impact and our sales are so sort of [14:56] sales are uh reflecting that as well and [14:58] growing very rapidly. [15:00] In terms of complexity, me personally, I [15:02] come from sort of L4 background. I have [15:03] worked my rest of my career on [15:05] autonomous cars. And I can really say [15:07] that the complexity of autonomous cars [15:09] is is and these drones is is really [15:11] similar. We have a tremendous amount of [15:13] technology that is really fit in a very [15:15] tiny body that flies above cities, you [15:18] know, above above densely populated [15:20] areas in a very high speed. So, in terms [15:23] of the amount of technology that goes [15:24] into these little things, it it it is [15:26] tremendous. Everything from, you know, [15:28] starting with autonomy, just making that [15:30] thing fly, fly reliably in all of the [15:32] all of the scenarios, but all going all [15:34] the way to like making sure you have [15:36] reliable link to the drones, as well as, [15:38] you know, these are aerospace that are [15:39] sort of [15:40] flying systems at high speed. So, all [15:42] the aerodynamics need to be considered. [15:44] There is like there's a lot of [15:45] technology that is going into very small [15:48] systems. [15:49] So, who do we serve? We really serve [15:52] really [15:53] the backbone of our civil civilization. [15:55] And we have we have demonstrated at this [15:57] point of time that these these flying [15:59] robots have life-saving impact to all of [16:01] these all of these industries. [16:04] What we are seeing at H transition in is [16:07] that it's not anymore these systems are [16:09] toys. If these systems became an [16:11] infrastructure, meaning that these [16:13] systems must be always available 24/7, [16:15] 365 days a year, and they need to be [16:18] always functioning when a pilot needs [16:20] them, they need to work, and they need [16:22] to fly there. So, it's really like we we [16:23] have transitioned from the era in early [16:25] 2000 when these things were just a toy [16:28] with sort of some camera on it that you [16:29] could make a cool shots and a cool [16:31] videos to something that people rely on [16:34] for a day daily work and save lives [16:36] across all our industry across many [16:38] industries. [16:39] And the AFR is the first example where [16:41] flying robots flying robots really [16:44] scaled and became an infrastructure. [16:46] This point, they are they are deployed [16:48] across the entire US. [16:50] And we are seeing this scale also [16:52] reflect in our annual as fly rate. So, [16:54] this is how many flights we are doing [16:56] per year. And you know, we started in [16:58] 2021 with sort of small numbers, but we [17:00] are seeing this huge spike which was [17:01] mainly driven by, you know, our [17:03] introduction of X1, which is our main [17:05] main workhorse. But, you know, truly, if [17:08] I even if I even further dig down to the [17:11] Doc Plus X1, that's truly the key [17:13] enabler that at this point of time these [17:15] systems are deployed completely [17:17] autonomous, always ready for anyone that [17:19] need them in the most critical time, and [17:21] just and just by as I as I did in the [17:22] beginning, I can just uh press a button [17:24] and they take off and and and fly. [17:28] Uh [17:29] more than we have DFR flights, so these [17:31] police officers are flying more than 1 [17:33] million flights per year. [17:35] But if you think about it in a in a in a [17:36] in a global or or on a larger scale, [17:38] this is really a tiny fraction of all [17:40] 911 calls. And our goal is really that [17:43] the drone is always going to respond [17:44] first to 911 call because it can arrive [17:46] there faster as we have shown by metrics [17:48] by many of our customers, and as a [17:50] result sort of can can give this [17:52] critical information much earlier than [17:54] than before. [17:55] We are we are uh the docks are close to [17:58] 610 [17:59] uh distribution poles right now. But [18:01] again, if you take it in the even across [18:03] just across US, there's 180 million [18:06] distribution poles that we can always [18:07] keep up to date, inspect, and make sure [18:10] there is no damage. So really [18:12] and on on on on the on a SRR side, we [18:16] have shipped we have delivered more than [18:17] around 3,000 X Tendees to the [18:20] uh to the SRR contract, but if you think [18:22] about it, that's just like 1 week of [18:24] usage across Ukraine. So if you if you [18:26] think about it at a at a global global [18:28] scale, we are just starting. This is [18:31] just the beginning. Just the beginning [18:33] and the impact that we are seeing that [18:34] we can make, the benefit we deliver, and [18:37] the opportunity is just is just [18:38] enormous. [18:40] So but truly like if you are thinking [18:42] about, you know, responding to 200 10 [18:45] calls, can can this this overall concept [18:47] that we have right now scale? And really [18:49] what's what's sort of the outcome of the [18:51] the autonomous team mission that's [18:52] Skydio? [18:53] So if I if I play forward the [18:54] simulation, this is a Clark County [18:56] simulation where we have [18:58] around 500,000 calls in our data set. [19:01] And if I [19:02] obviously, if you are thinking about [19:04] where you want to strategically deploy [19:06] your docks, you first want to deploy the [19:08] docks in the most sort of densely [19:10] populated areas where you are seeing the [19:11] most 911 calls. So, you start there, you [19:14] start with 15 dogs, you have 15 [19:16] operators, but as you can sort of [19:17] continuously scale it up and you want to [19:19] cover all of them, it becomes 160 dogs [19:22] requiring 160 humans to fly them if you [19:24] want to do one-to-one pairing. And [19:26] obviously, one thing that we are hearing [19:28] very very [19:29] very often is that finding, you know, [19:31] high-quality pilots and and training [19:34] them to be very proficient in their in [19:36] in their in their duties is extremely [19:38] extremely difficult. And that's one of [19:40] the sort of one of the key thing that we [19:42] have to overcome to enable this overall [19:44] scale. We have introduced a capability [19:46] and I've showed you to earlier that, you [19:47] know, you can go from one-to-one, so [19:49] meaning one-to-one pairing that one [19:51] pilot operates one drone, to something [19:53] that one pilot can operate at least four [19:55] drones at the same time. [19:57] But even then, [19:58] um [19:59] even then, that just cuts to 42 42 [20:02] operators. Our goal is really to go to [20:04] as close to zero as possible, meaning [20:07] that pilots are going to become this [20:08] sort of high-level orchestrators that [20:10] they can operate their own fleets just [20:12] through high-level commands and and task [20:14] them with high-level high-level [20:15] executive executive missions that these [20:18] drones can sort of together perform and [20:20] and solve. [20:22] Um so, we're giving pilots superpowers [20:24] and we haven't done it [20:26] we of course don't do it, you know, from [20:28] zero to 100 immediately. We are all [20:30] engineers, so we sort of distribute the [20:32] problems to smaller pieces to solve them [20:33] incrementally. So, this is one of the [20:35] features that, you know, where we [20:36] started [20:37] uh where we where we gave people [20:39] superpowers that they can navigate from [20:41] point A to point B even in some [20:42] something like dense dense urban [20:44] environment. This is a San Francisco. [20:46] So, all of the people have to do is just [20:48] like all pilots have to do, fly there [20:50] across the city and the drone does [20:52] everything completely autonomous and the [20:54] the the human doesn't have to really [20:55] think about it. So, that's really one of [20:57] the areas that you just task the drone [20:58] to go there and it performs everything [21:01] autonomously. So, how does it work under [21:03] the hood? What technology does it really [21:04] rely on? [21:05] Uh so, we do rely on maps. We have [21:08] multiple source of the maps. We use a [21:10] digital surface map, digital terrain [21:11] map, as well as more of a structured [21:13] information where for example our power [21:14] lines. We combine all that information [21:17] into one sort of map representation that [21:19] then we sort of then we load to the [21:21] drone. And on top of it, you know, once [21:23] you have the map, once you know roughly [21:25] where the buildings are, then you can [21:26] use graph search or you know, some [21:29] uh, some methods like that to find you [21:31] know, what's the most efficient most [21:33] efficient path to travel from point to [21:35] point A to point B fairly efficiently. [21:38] But I'm sure like all of you have worked [21:40] on robotics and and all of you know that [21:41] anything that relies on a map has to a [21:44] fundamental problem. Maps are never up [21:46] to date. You know, that's that's that's [21:48] a fundamental problem that we have to we [21:50] have to overcome as well. So in order I [21:53] am and obviously if if you're flying [21:55] through a city and you don't have a [21:56] building in your in your map that that [21:57] that can cause a lot of problems. So [21:59] what we do [22:00] is that as as as customers are flying [22:02] throughout throughout throughout the [22:04] cities we we take all the data, we do [22:07] dense bundle adjustment and dense [22:08] reconstruction, we reconstruct all of [22:10] the areas as the as the drone flies [22:12] thanks to our top and bottom nav [22:13] cameras, and [22:15] we can then identify and compare the two [22:17] maps against each other and see whether [22:19] there are areas that you know, that we [22:20] have reconstructed but are not in the [22:22] map and identify exactly as in this case [22:25] this giant this large crane and and [22:27] whole sort of foundations of a building [22:29] uh, of a building and we can update the [22:31] map update the map with additional [22:32] information. And as you can imagine, you [22:35] know, if you're going in Miami or San [22:36] Francisco or the biggest cities, there's [22:37] always construction. So this turns out [22:39] to be very very critical part to really [22:41] make it make sure that this this this [22:43] system can work always as I mentioned [22:45] and and and and always reliable. [22:49] Um, [22:50] yeah, and the the fact we are flying in [22:53] dense urban environment. And as you as [22:54] you know, if you go with your phone down [22:56] in San Francisco, the GPS on the phone [22:59] it becomes very unreliable. Yeah? [23:02] >> You guys have the answer for the digital [23:03] twin of the city you guys have? [23:06] We generally we generally get various [23:08] sources of the map. You can think of it [23:10] the maps can be generated from [23:11] satellites or or whatever. We can we [23:13] first start there and then we build on [23:15] top of it. [23:19] Uh I mean depending what you mean by [23:21] full digital because we don't [23:22] necessarily need in this particular case [23:24] about how it looks or how pretty it is. [23:26] We mainly care about there is a building [23:28] and may may mainly about more like [23:30] occupancy map of some sort. So that's [23:32] the main thing that we rely on. [23:35] I mean it is 3D. It is 3D. So yes. [23:38] From that perspective. [23:41] Uh and then if you have any questions of [23:42] course always ask. Please. Um [23:45] So yeah. Um global pose estimate. Uh as [23:48] you know if you go down down San [23:49] Francisco in the city GPS tends to be [23:52] very unreliable. The main reason why [23:54] that actually happens is because the [23:55] satellites are actually occluding. There [23:57] is multipath effect that you know the [23:59] the the that [24:01] makes the measurements unreliable. But [24:02] because we have access to the you know [24:05] the or or all sensor measurements on the [24:07] drone as well as the camera information, [24:09] we can fuse both of them together and [24:11] smartly you know see where the [24:13] satellites are actually on the in the [24:14] sky and then reject them which we know [24:17] in the image space are occluded. And as [24:19] a result we can actually build a sort of [24:21] GP GPS pose estimate that is [24:23] substantially more reliable and accurate [24:25] in in exactly these scenarios where we [24:27] care the most about. [24:30] Um the other one is shadow. [24:33] Uh as for the DFR, you know, you can [24:35] imagine that if you fly to a location [24:37] very often there is like you are looking [24:38] for something or or someone and then you [24:41] want to once you find them you want to [24:42] lock on them and make sure you always [24:43] keep the kill the keep them in field of [24:45] view. There's another area of very very [24:47] active work for us where we want to [24:49] where we want to get to the state that [24:50] you actually just lock on a sub lock on [24:52] a subject uh whether it's a car or a [24:54] person and uh whether it's during the [24:56] day or night and you really don't have [24:58] to care about anything else. That's [25:00] That's a really fundamental capability [25:01] that we are investing a lot of a lot of [25:03] time and an effort to. [25:04] Um how it works under the hood is that [25:06] we have, of course, we have a tracking [25:08] model [25:09] that is able to identify for us a [25:11] location of the subject in the in the 2D [25:14] sort of image space and then project it [25:16] based on the location of the drone on a [25:18] map and get a 3D coordinate. [25:20] Uh the one thing that we have introduced [25:22] uh well uh [25:25] Hopefully it's going to go. [25:26] Yeah, one thing that we have seen one [25:28] thing we have seen when we deploy this [25:29] in the real world and then again people [25:31] start using in real applications is that [25:33] if you track a car, obviously drone is [25:35] not as fast as a car, and very often you [25:37] fall behind. And in that case as in and [25:40] in that scenarios as as you can see [25:41] here, you're we are trying to the person [25:43] is trying to track this car, very often [25:45] this car goes behind a building. Once [25:47] the car goes behind a building, you have [25:49] no idea where the where the car or or [25:50] subject of your interest is, you lose [25:53] you lose line of sight, and you don't [25:54] know what's going to happen next. [25:57] The way we have solved it uh is that we [25:59] have trained we have a trained a deep [26:00] learning model [26:02] where we where we take actually the you [26:03] know the same map information that I've [26:05] told you about that we that we care a [26:07] lot about that is curated as accurate as [26:09] possible, and we feed it through uh [26:12] through a neural network to actually [26:13] output directly the actions. And that's [26:15] really our next generation That's our [26:17] next generation of a shadow that we have [26:19] introduced [26:21] on Ascent. The way we the way we train [26:23] this policy is it has multiple stages. [26:26] So, first of all, you know, the encoder [26:27] is of course uh much larger network than [26:29] the than the action policy itself. So, [26:31] we pre-train the policy based on, you [26:33] know, making sure that predicts [26:34] visibility map because that's actually [26:36] what you care about. You want to know [26:37] where the subject where you have to be [26:39] so that you have visibility of the map [26:41] in addition to other information. We [26:43] take the sort of What we do is that we [26:44] take the larger map and we can encode in [26:46] an in a in a in a embedding in much [26:49] smaller dimension such that you [26:50] afterwards we can take this embedding [26:53] and [26:54] train it [26:55] uh [26:57] train it in RL fashion at a at a very [26:59] large scale. So we we use simulation for [27:02] this [27:02] where we deployed it in where where [27:04] where we deployed it in start with San [27:06] Francisco. We start with San San [27:08] Francisco. We train the policy there and [27:09] of course over time we want to make sure [27:11] that this works across the entire [27:12] country and all the cities. So we scale [27:14] it substantially across all the maps [27:16] that we that we have and [27:19] and afterwards uh we deploy it on the [27:21] drone. [27:22] This is another example that uh there [27:25] very interesting behaviors that fall out [27:26] of this policy. This is an example in [27:28] San Francisco where the car actually uh [27:30] came to came to a stop and went to the [27:32] right and we have seen that actually the [27:34] model learned that before as the car [27:36] comes to a stop it has to actually speed [27:38] up and go right on top of it such that [27:40] it has such that it is positioned in the [27:42] best possible way [27:44] uh whatever the whatever the drone [27:46] whatever the car does next. So it has [27:48] the best chance actually to keep the [27:50] visibility. Yes. [27:51] >> This is very impressive. I'm curious [27:53] what's the fallback in this. [27:55] For example like Are you mentioning in [27:57] the other building [27:58] >> Yeah. [28:00] Yeah in our in our case what it what [28:01] would happen we would lose the we would [28:03] lose the subject because we don't know [28:04] what happens and in that case [28:06] uh the human has to take over. That's [28:08] the right that's the current situation. [28:10] Yes. Yes. We would flag it and the human [28:12] has to has to take over. [28:14] Uh I will [28:17] go back to this video [28:20] where Yeah this was actually in an hour [28:22] ascent our yearly conference that [28:23] happened few months ago in in in uh [28:26] actually month ago in in Vegas where the [28:28] car actually came came to make turn [28:30] around the garage and even here you can [28:32] see it's going to turn right again and [28:34] the policy is going to actually decide [28:35] to swing around the building much [28:37] earlier actually and anticipate the car [28:39] actually will do that such that it is [28:41] positioned in the best best location [28:43] such that it has field of view uh line [28:45] of sight to that uh to that subject. [28:49] Um [28:51] Yeah, so that's I talked a lot about [28:52] DFR, but now I want to switch gears a [28:54] little bit to talk about inspecting and [28:56] inspecting critical infrastructure. [28:58] So, the way the way the software [29:00] generally the way that the software [29:01] generally works is that you can select [29:03] all the assets that you load all the [29:04] assets into the software, and then you [29:06] can select which assets you want to [29:08] inspect, and the software automatically [29:10] generates all what are all of the areas [29:13] or what are the actually shot shots that [29:15] I have to take of the assets such that I [29:16] have the best coverage of the things I [29:18] care about the most that then can be [29:20] inspected. [29:21] All what you do is just a select which [29:22] of them you care about, uh [29:24] pick what type of shot sheet as we call [29:26] it, uh [29:27] uh and and take the and and execute the [29:29] mission in completely autonomous manner. [29:31] Then we have on these drones [29:34] um [29:35] on accident we have RTK attachment that [29:37] you can actually just mount on it, which [29:38] gives you as you all know very reliable [29:40] pose estimate. So, then we can just go [29:42] and take all of these shots of these [29:44] assets that are then loaded into the [29:46] system and and a human can sort of [29:48] inspect them to figure out any possible [29:50] any possible anomalies or anything that [29:52] needs to that needs to be inspected. [29:55] Uh [29:56] what it turns out when deploying these [29:57] things [29:58] that [30:00] uh as you can imagine what you're [30:01] relying on is the one thing uh which is [30:03] the [30:05] location of the assets. [30:06] And that turns out to be a big problem. [30:08] Many of the utility customers some of [30:10] the utilities customers have actually [30:11] very precise location of these assets [30:13] because they have spent the time and [30:15] energy and and money to actually create [30:17] high precision lidar point clouds of the [30:19] entire entire grid, but it's not really [30:22] it's not really the case across all the [30:24] all the utility customers out there. [30:27] So, the thing that we have to build is [30:29] in order to make this work really like [30:31] how do you get actually the precise [30:32] location of the assets? [30:34] So, the the way that these things are [30:35] actually deployed right now [30:38] uh [30:40] uh is that you first do a mapping [30:41] mission around these assets that needs [30:43] to be sort of roughly around around the [30:45] location of the assets. You know [30:46] generally at the beginning like this is [30:48] roughly these assets are in those [30:49] locations. Then you detect them, you [30:51] sort of the you do you do you detect [30:53] these assets in the in the in the in the [30:55] camera field of view and we do the [30:56] complete 3D reconstruction of the scene. [30:59] And once we have that, we can easily [31:00] identify actually the the the the asset [31:03] and then match them to their original [31:04] prior guess where they are. Which these [31:07] guesses are sort of sometimes very [31:08] close, sometimes further, but because we [31:10] have them because we already know all of [31:12] them, we can sort of do the best [31:13] matching to to their prior guesses. So [31:15] yeah, we built as I mentioned it's very [31:17] important that we built sort of the [31:19] the entire dense point cloud and [31:22] afterwards and I once we have a dense [31:24] point cloud and we also have a [31:25] detections of these assets, it is fairly [31:27] straightforward to figure out which of [31:29] the points [31:30] uh, lay [31:32] uh, which of the points are on the [31:34] assets and once you have that you can [31:36] figure out exactly their 3D geometry, [31:38] can figure out their the what's what's [31:40] their what's their top point and [31:41] afterwards it is fairly straightforward [31:43] to figure out what are the shot sheets, [31:44] what are locations of the photos that [31:46] you have to take so that you cover as [31:48] much of the assets as possible or the or [31:50] the areas of the assets um, [31:51] that that you need. [31:55] Uh, [31:57] yeah. [32:00] Uh, okay, I'll I'll move forward. Um, [32:03] the [32:04] so that was inspection and now the last [32:05] one which is perhaps the newest new the [32:07] newest for us, which is securing [32:09] critical infrastructure, securing [32:11] securing areas. Yeah, this is fairly uh, [32:14] new area for us. [32:16] Uh, [32:18] where the main goal is that you have [32:20] many of these docking stations spread [32:22] across your, you know, it could be for [32:23] example Stanford universities or or or [32:26] sort of headquarters or or utility [32:29] customers and you can plan a mission [32:32] that you want to perform. You want these [32:33] systems to fly completely autonomously [32:36] and then all that they do is they fly [32:37] 24/7 all the time on these areas. And [32:41] then you can draw on the on the air on [32:42] the on the ground these sort of [32:44] detection areas that you care about, [32:45] meaning I really don't want that there [32:47] is anyone going to be in this area. And [32:49] if there is someone, I really want to [32:51] know about it. [32:52] We have, as I as I showed you before, we [32:54] have already sort of the people and car [32:56] detection system running on or running [32:57] on board of the drone, and we know where [33:00] these detections actually happen. So, [33:02] every time if we detect people or car in [33:04] these air in these detection areas, we [33:07] can just send an alert to pilots or or [33:09] operators that hey, there's someone in [33:11] these areas, pay attention to it. That [33:13] means that really means that what what [33:15] happens is that you can have one pilot [33:16] have, you know, operating the fleets of [33:18] the drones across completely different [33:20] areas occurring the country, have these [33:22] systems operating in autonomous manner, [33:24] and all what you do is you pay attention [33:26] to if an alert happens. [33:28] As you can imagine, the one very very [33:30] important thing here is the reliability [33:31] of these models. Every model that we [33:33] have to train for this application needs [33:35] to be highly accurate, because if you [33:37] imagine that you have, you know, tens or [33:38] or hundreds perhaps drones across the [33:40] entire across the entire country or [33:42] across your deployment, if you were to [33:44] getting alerts all the time, you would [33:46] never be able to keep up with it. So, [33:47] the false positive rate really matters [33:49] here. [33:51] Uh but what we have noticed that [33:53] uh what we have noticed [33:56] uh that depending on the use case and [33:59] application uh sometimes just people or [34:03] cars is is not enough. Sometimes you [34:05] want something, you know, more unique [34:07] for your use case. So, really the [34:08] direction that we are going for going [34:10] towards next is work on the models that [34:12] are more going direction of open [34:14] vocabulary, so you can really type [34:16] whatever whatever is your very interest. [34:18] And whereas in this case is sort of [34:19] wooden wooden pallets. And if you if you [34:22] see the wooden pallets in the in the in [34:23] the field of view, you get these alerts, [34:25] and you notify the pilots to actually, [34:27] you know, pay attention to and and take [34:29] over and figure out what's actually [34:30] happening. In this case, I was looking [34:32] for wooden pallets and I think later [34:34] there's also going to be doors. But [34:36] really the the the direction that we are [34:37] going towards is to is to build these [34:39] models in such a way that are as [34:41] generalizable as possible [34:43] leveraging the [34:44] uh the latest foundation models. [34:48] But you can ask like, "Hey, uh these are [34:50] big models, right? If you are thinking [34:51] about open vocabulary or anything text [34:53] specific, these are, you know, hundreds [34:55] of millions going toward multiple [34:57] billions of parameters models. Can you [34:59] run on the board?" On NX 10, we have we [35:02] have NVIDIA Orin NX. So, these are [35:04] fairly small compact computer, fairly [35:06] power efficient. So, obviously, we [35:08] cannot run it we can we cannot run it on [35:10] the on the board. The way we actually [35:12] solve this is we run all we run always [35:14] in the cloud. One one very critical [35:16] piece of piece for us is [35:18] the most important thing that these [35:20] systems are are giving to customers is [35:22] the real-time sort of video that we [35:24] could actually can look at where the [35:26] drone is, and then we can we always of [35:27] course stream [35:29] stream the footage from the board from [35:31] the drone through the cloud to the [35:32] everyone to the everyone's computer. [35:34] And uh only only way that is actually [35:37] So, this is very very important [35:38] fundamental part of our architecture and [35:41] our our product. So, we are investing a [35:42] lot of resources and time into it. The [35:44] way it works under the hood and really [35:46] what enables it that we have multiple [35:48] source of connections. We have [35:49] point-to-point point-to-point link. So, [35:52] every docking station comes with sort of [35:53] external radio that, you know, can talk [35:55] to the drone. So, we have sky we have we [35:57] have sky dealing as we call it. But at [35:59] the same time, in many of our [36:01] deployments, these drones have SIM cards [36:03] in them and so we have 5G. And you know, [36:06] we don't have just one, we have [36:07] multiple. So, we are really fusing all [36:09] of these sources, all of these [36:10] connections together into one sort of [36:12] connect fusion that allows us to have [36:15] always low latency and high resolution [36:17] video always available. And that is [36:19] really fundamental piece that allows us [36:21] to take these train these very very [36:23] large models and and deploy them in the [36:26] cloud. [36:27] And and always always work. [36:30] Um [36:31] obviously, you know, there are various [36:33] various uh [36:35] there are various tasks. Every every [36:37] customer, every every use case may [36:39] require different type of drone and [36:40] different type of vehicle. [36:42] Uh we have uh recently launched [36:44] something new drone called uh Skydio [36:47] R10, which is a smaller form factor that [36:49] has sort of prop guard around it and can [36:51] actually go in the very tight areas. [36:53] Uh that is used across uh SWAT [36:56] deployments or utility inspections. You [36:58] know, it flies in a very tight tunnels. [37:00] The very very tight um tunnels or areas. [37:03] It has still obstacle avoidance. It has [37:05] two fisheye cameras on board. So, we [37:07] still provide sort of a front-facing [37:09] obstacle avoidance capability. So, you [37:12] can still go, you know, full full speed [37:14] through a through a through a small [37:16] tunnel or perhaps even go into the to [37:18] the inside of a building and not really [37:20] worry about crashing into anything. And [37:22] that turns out to be extremely [37:23] important. [37:25] Well, [37:26] >> [clears throat] [37:27] >> uh very often, you know, as I said, uh [37:29] when you are thinking about DFR [37:31] deployments, the one critical factor is [37:33] like you want to fly as fast as possible [37:35] from point A to point B, right? So, we [37:37] always get the question, "Hey, I want to [37:39] fly faster. And once I fly there, once I [37:41] get there, I want to be there as long as [37:42] possible." So, those are the the two [37:44] critical asks that we are getting we are [37:46] getting we are getting very often. [37:48] The way um the way the right way to [37:50] solve it actually is to look into, you [37:52] know, what is available out there in in [37:53] planes. [37:55] Planes can actually fly very fast. They [37:56] can actually stall stay in the air much [37:59] longer because they are much more [38:00] efficient. [38:01] But if you think about it, you know, if [38:02] you look out there, what are really are [38:05] there really any existing fixed-wing [38:07] aircraft that are available at this [38:08] scale and sort of are able to fly in [38:10] dense urban environments? [38:12] If you look at it, [38:13] not really. Um there are, you know, the [38:15] they are various of these systems that [38:16] require huge crew on the ground to catch [38:19] them, deploy them, and make sure that [38:20] they are always available, always ready [38:22] to sort of uh to to to fly. [38:25] And even once they are actually deployed [38:27] and they fly, even then you may need [38:29] even multiple pilots for a single drone [38:31] to be operating them. [38:33] And that that really doesn't, you know, [38:35] match nicely with, you know, what what [38:37] we what we are envisioning that you can [38:38] have fleet of these small drones that [38:40] you can operate [38:42] operate multiple perhaps at the same [38:43] time. [38:44] So, [38:46] if you look at a we looked at we looked [38:48] quite hard in the industry and tried to, [38:49] you know, see what what works out there [38:51] the best. So, this is one of the example [38:53] of fixed-wing vehicle that that that is [38:55] perhaps the best in the [38:57] was the best in the [38:59] in the in the in the category. [39:01] Uh and one thing that you can notice it [39:03] actually has the the the camera mounted [39:05] in the center of the fuselage. And if [39:07] you think about it, they often they [39:09] often think what you what you actually [39:10] want to do for our use cases is that you [39:13] want to fly somewhere and look at [39:14] something. Very, very basic, right? [39:16] And if you if you if you simulate it, [39:18] what you can actually see that if you [39:19] just fly somewhere look at something, [39:21] every single time you loiter around a [39:23] point, you will see that actually there [39:25] is a fit part of the wing that that [39:27] actually occludes your visibility or [39:29] your your actually subject that you're [39:31] looking at. And that means that if you [39:32] are looking at something that is very, [39:34] very important, uh [39:36] you will lose it. You will you will very [39:38] likely lose it, especially if you are, [39:39] you know, shadowing or tracking someone. [39:41] Uh that's fairly high frequency task. [39:43] So, if you lose them even for split of a [39:45] second, it's very likely you will lose [39:47] that subject. [39:48] So, the way we [39:51] uh [39:52] the way we [39:54] The So, the problems overall that we [39:55] have to solve is fir- first of all, we [39:57] want to make sure that we can automa- we [39:59] can do autonomous fully land and and [40:01] launch these ve- these aircrafts. You [40:04] want to make sure that the flying [40:05] experience is very straightforward. All [40:07] what you have to do is just a click at a [40:09] point look at like and and just the the [40:10] the vehicle looks there. [40:13] And we also always want to want to make [40:15] sure that you always [40:16] um maintain the visibility no matter [40:18] what. [40:20] To solve it, uh we have introduced um [40:23] new new product Skydio X10 that I have [40:26] also, you know, here for everyone to [40:28] show perhaps afterwards. It is a fairly [40:30] unique vehicle. [40:32] And I'll touch on a little bit later, [40:33] but the way it actually works that we [40:35] have this docking station we call [40:37] Megadock that is a [40:39] that's actually container and within the [40:41] container you have sort of a shelf where [40:43] multiple of these systems are charging [40:45] and living as well as a robotic arm. And [40:48] that's a that's a fun part. All of the [40:50] robotic arm does is actually just catch [40:51] it like takes the one of these one of [40:53] these vehicles from the shelf, launches [40:55] it, and once it once the vehicle comes [40:57] back, sort of picks it up again and puts [40:59] it in the shelf. And you can then as a [41:01] result, you can operate multiple of [41:02] these vehicles just from a single dock. [41:05] And [41:06] because this is sort of a [41:08] this is a vehicle that, you know, it's a [41:09] sort of a tailsitter, it can sort of go [41:12] up from from the docking station and [41:13] then just transition to forward flying [41:16] to to to forward flying which is [41:17] substantially more efficient. And as a [41:19] result, you have the nice thing of [41:20] everything. You are able to auto [41:22] completely autonomously take off and [41:24] land. You are able to [41:26] and and uh you you have very efficient [41:29] systems. All the complexity of actually [41:32] taking off and landing is on the docking [41:34] station, so it's on the ground. It never [41:36] flies. And as a result, we can have we [41:39] can have completely we can have very [41:41] efficient very efficient vehicle that [41:43] just flies. [41:45] It doesn't need to carry this additional [41:47] weight all the time. [41:49] So, that solves the first part, making [41:51] sure that we can always sort of take off [41:53] and land without any without any [41:55] additional people. But the other thing [41:57] is that I want to put I want to, you [41:58] know, point everyone's attention to [42:00] is nicely asymmetric, right? It's not [42:01] really as a unique is not really the [42:03] same as as the conventional vehicles [42:05] that are perfectly symmetric. You We [42:07] have We mounted the gimbal on the left [42:08] side. I'll try to show it maybe this [42:10] way. On on on end of the wing so that if [42:13] it sort of orbits you always, you know, [42:14] retain the line of sight to the point [42:16] that to the point that uh if you look at [42:19] somewhere over there. [42:20] And as a result, you know, we had to of [42:22] course counterbalance the fuselage more [42:25] on the right side. So, it's it as a [42:26] result, you know, it's it's perfectly [42:28] asymmetric. We have two control we have [42:30] control surfaces we have two flaps and [42:32] of course two motors. [42:34] And in addition, we have two nav cams in [42:35] the front so you can always we can we [42:37] have also obstacle avoidance. [42:39] Uh from our similar to our [42:42] uh extend. [42:46] And yeah, as I mentioned, you know, the [42:48] fundamental break fundamental [42:49] breakthrough here is that the gimbal is [42:51] in this case mounted on the left side on [42:53] on left uh left uh [42:55] on left wing at the end so we have [42:57] always perfect visibility and we never [42:58] have blind spot. [43:00] As a result, you know, the other thing [43:01] so we have solved now we don't have any [43:03] occlusions. We can take off and land [43:05] now. We have to fly the way easily. So, [43:07] we have developed a [43:09] a new autonomy or or a new autonomy [43:11] system for [43:12] for this vehicle which is sort of [43:14] relying on optimization optimization [43:15] based methods. The all what you have to [43:17] do is [43:18] look at a look at a like you have to [43:20] just say that I want to look at that [43:21] point and that that higher level planner [43:23] just figures out exactly what is the [43:24] most optimal trajectory that I have to [43:26] take while sort of of course making sure [43:29] that this always flies the most [43:30] efficient most efficient region for this [43:32] vehicle is always forward flight so you [43:34] never want to de-transition. So, you [43:35] loiter. It's also prior such that you [43:38] always loiter more to the left side so [43:40] you never so you never actually, you [43:42] know, uh are are going to the right side [43:44] because that then you may lose line of [43:45] sight. [43:46] And that's really what what what's the [43:48] final thing that we had to solve is [43:49] enable that you can just set a look at [43:52] point and and that's all what you have [43:54] to do. You don't have to do [43:55] you don't have to worry about where the [43:57] vehicle flies. It does everything [43:58] completely autonomously. [44:01] Uh [44:03] and you know, that that was mainly on [44:05] the planning front but you know, [44:06] fundamentally underneath there is, you [44:08] know, we are building on on on fairly [44:10] matured stack that, you know, we are [44:11] using our VIO system from extend. So, we [44:14] are using still our nav cams that are [44:16] there on on on the belly to figure out [44:18] the pose estimate of the vehicle, as [44:20] well as, you know, when we are when we [44:21] are going to land to a dock, we still [44:22] detect the checkerboards that are [44:24] similarly on our on our docking station [44:27] to figure out the relative [44:28] transformation between the dock [44:30] between the arm and the vehicle and then [44:32] control the actually control the arm so [44:34] that you can just traps the vehicle as [44:36] it's landing. [44:39] Uh [44:41] So, [44:42] you know, if you if you take step back, [44:43] I do think that we have, you know, we [44:45] have we have [44:47] very matured autonomy stack that is able [44:49] to able to control various vehicles [44:52] across our all entire portfolio. So, [44:54] what is the fundamental key enabler that [44:57] that allows all of this? [44:58] And I think it's really this loop. [45:00] >> [snorts] [45:00] >> The [45:02] I think this loop, like how quickly you [45:03] can actually deploy stuff to the real [45:05] world, get the data back, and you know, [45:07] retrain your models and evaluate them, [45:09] and how quickly you can turn this this [45:11] this [45:12] um [45:13] this crank really determines how quickly [45:14] you can move. At Skydio, you know, I [45:16] think we have at this point the largest [45:18] sort of data set of of flying robots out [45:20] there, and we are of course partnering [45:22] with some of the some of our customers [45:24] who are who who choose to give us access [45:25] to their data that we of course [45:27] aggregate and de-identify and all do all [45:29] of those things [45:30] to build sort of huge data set that we [45:32] can then train all of these models on at [45:34] large scale and really build the most [45:36] capable autonomy stack that is out [45:37] there. [45:39] Uh so, what's next? Uh well, more [45:42] robots, obviously. We really love flying [45:45] robots, and you can definitely expect [45:46] that over, you know, the next over over [45:48] over the next few [45:50] over the next time, we are going to see [45:51] more and more robots actually flying out [45:53] there. [45:54] And but the other thing is I do think [45:56] that one next step that for us that [45:58] really is going to enable and, you know, [45:59] perhaps everyone in the research is [46:02] uh that is going to enable this [46:03] transition from from a drone being sort [46:05] of something that you have to task in [46:07] intermediate levels is that if the if [46:09] the drone becomes your body [46:10] um and you spoke to it just with natural [46:12] language. And that's really the [46:14] direction that we are heading on heading [46:16] towards is to enable that you can just [46:17] you can just coordinate these drones at [46:19] very high level commands and they they [46:21] perform these task task and entire [46:23] missions completely autonomously. [46:26] With that, [46:27] um [46:30] uh [46:33] Yeah, uh with that I think that's [46:35] everything from my front. [46:36] Thank you very much for listening. [46:38] Uh in general, uh we are of course [46:40] hiring across across our entire [46:42] portfolio in terms of what company we [46:44] are. You know, we are as I mentioned at [46:46] the beginning, we are building the [46:47] hardware, the fundamentals of like all [46:49] of our all of our all of our robots. We [46:51] build the software, the low level [46:52] software, the high level software as [46:53] well as the cloud. And you know, this [46:55] this requires the top talent across all [46:58] areas to come together so that we can [46:59] optimize the entire stack end-to-end. [47:02] So, if if if flying robots are something [47:04] that excites you, please come talk to [47:06] talk to myself or Paul down there in the [47:08] middle. Uh and and [47:11] we are happy to to chat on. We are [47:13] hiring across all the fronts. Thank you. [47:16] >> [applause] [47:19] >> Cool. [47:22] Do you have any questions? Happy? Yeah, [47:24] right there. [47:25] >> I have two questions. [47:26] >> Yes. [47:27] >> The first use case that [47:30] you talked about early smart use cases [47:32] was how do you scale operations for [47:34] light on the ground response? [47:35] >> Yeah. [47:36] >> Um [47:38] you have operators who pick up the [47:40] calls. [47:40] >> Yes. [47:41] >> Are those the same people who control [47:43] the drones? If they're not, what [47:45] capabilities [47:47] does [47:48] your PD have to build up [47:50] for them to scale up to the [47:53] deploying these [47:55] these drones at a scale of the city? [47:57] >> Right. [48:08] Yeah, great great question. So, the [48:09] question was what the what is generally [48:11] the workflow for PDs with if they got a [48:13] 911 call and um [48:16] whether they are they are the same [48:17] people that actually controls these [48:19] drones. So, no, they are they are not [48:20] the same people. Generally, these calls [48:21] arrive in the 911 911 centers then they [48:24] relay to the police officers to actually [48:26] the pilot the drone. So, those are very [48:28] those are different people. However, [48:29] there are there are companies that we [48:31] are working with such as Axon that sort [48:33] of has a solutions that sort of [48:35] transcribes right away these these 911 [48:38] calls into a text that then feeds into [48:41] our system as a markers. And actually, [48:43] we also have a we have also a system [48:45] that automatically priors these drones. [48:47] So, once the 911 call is transcribed, we [48:50] get the markers, we look at the what [48:52] what the what the what's the call what's [48:53] the call about and [48:55] we determine whether we should actually [48:57] based on a policy that every single [48:58] customer chooses themself whether they [49:00] want to dispatch to such a call and then [49:02] a drone is automatically open and [49:04] actually are sort of prime to sort of [49:06] just lounge. And all of the people then [49:07] have to do or pilots is just click [49:09] lounge and they take off autonomously. [49:11] So, that sort of there exists already [49:12] solutions and we have actually a [49:15] capability that that that that we have [49:17] announced at a send this year that that [49:19] allows all of this. [49:21] Cool. [49:22] What does it [49:25] Yes. [49:31] Yes. [49:37] Uh [49:38] yeah, so no, these these these pilots [49:41] for now for for DFR are not part of [49:43] Skydio. These are generally officers [49:45] that the police department or whatever [49:47] customers of employees themself trains [49:49] themself. Every every [49:50] every single PD has their own policies, [49:52] you know, how when they actually should [49:54] operate a drone, so they go through the [49:55] training, but they are they are they are [49:57] not employed by Skydio. [49:58] >> The motivation behind the question was [50:00] how do you see the same as a company [50:02] from something like what happened in [50:05] >> Right. [50:08] Yeah. [50:10] Yeah, I mean it's every single like I [50:12] think what's the most important for many [50:14] of these deployments is that every [50:16] single you know police departments [50:17] defines their policy how exactly they [50:20] use these drones when they use and they [50:22] also sort of manage it themselves manage [50:25] it manage it [50:27] themselves, but one thing that perhaps [50:29] is the most important is transparency. [50:31] And one of the very critical feature [50:32] that we have that we have invested a lot [50:34] into and we have also you know announced [50:36] in in in during a cent is a transparency [50:38] dashboard. That means that every single [50:40] PD actually publishes every single [50:42] flight they do on this dashboard, so you [50:45] or anyone who is interested how these [50:46] drones are being used, they can actually [50:48] go and look at every time they fly. And [50:50] that's really I think the fundamental [50:52] thing for this technology in general for [50:53] the for for the program is is [50:55] transparency and make sure that is very [50:57] obvious [50:58] everyone can see and sort of perhaps [51:00] look at how these systems are being [51:01] used. And that's really the part that we [51:04] that that we pay a lot of attention to [51:06] as well. [51:07] Of course. Yes. [51:08] >> Do you have a partnership with the [51:11] companies with Amazon or other companies [51:13] for the [51:15] >> Yeah. [51:16] Yeah. [51:17] >> Is it like a in the future what your [51:19] business model [51:20] >> Uh so we don't do deliveries. That's not [51:23] really the business market we are in. [51:26] The the way that we work with Amazon and [51:28] with with these other these other [51:30] companies as as Zipline for example is [51:32] to make sure that these systems never [51:33] crash into each other. You know, that's [51:35] the part that we care a lot a lot about, [51:37] so we have direct collaboration with [51:38] them such that there is information [51:40] sharing and you know, when we fly [51:42] generally these are from 911 calls, so [51:43] they have higher priority. So they they [51:45] can sort of adjust their path in such a [51:47] way they never collide with with us. [51:51] Yeah. [51:52] >> Yeah, [51:53] I feel like every year I come to like [51:54] Skydio talk and [51:56] talk to people and it's great to see [51:58] your progress. [52:00] >> Cool. [52:01] >> My question is [52:03] if you try to indoor navigation [52:06] and the iOS main bottleneck [52:09] you know in the dark it's really kind of [52:11] features and you lose [52:13] because indoor GPS [52:15] would do better and then you know a lot [52:17] of people started putting lighter on top [52:18] is that something you consider or not? [52:20] >> So actually uh our drones actually [52:24] operate even also in in dark [52:26] environments and they have also obstacle [52:28] avoidance. The way we have solved it you [52:29] can mount and something attachment [52:31] called night sense. So it's infrared on [52:33] top and bottom that sort of illuminate [52:35] illuminates the environment to to [52:36] certain degree is not as perform as [52:38] during normal light and we have infrared [52:40] cameras. So that's a way that generally [52:42] you can you can solve it today. [52:43] >> Okay, maybe I was confused because you [52:45] were talking about the drone indoor and [52:47] then you had two fish eyes and so that [52:49] small [52:50] drone can also have that. [52:52] >> The the the small indoor drone also have [52:54] actually light. [52:55] >> Okay. [52:55] >> It also has light on it. The small one. [52:57] The the X10 in generally for X10 you [52:59] would mount the attachments and the [53:00] attachments can all be either infrared [53:02] or also normal light. Normal natural [53:04] light. [53:04] >> Got you. [53:06] >> Yeah. [53:07] >> Skydio you were talking about how you [53:09] want to reduce the number of operators [53:11] at work. [53:12] >> Yeah. [53:13] >> Um does that involve any use like [53:15] coordination swarming technologies by [53:17] Skydio? [53:18] >> Are you guys working on that? [53:19] >> How special [53:22] >> Yeah, so it's a very [53:23] >> So I mean these these drones generally [53:25] like don't they never communicate [53:27] between each other. Um so for these [53:29] especially for the public safety use [53:31] cases they always go through the through [53:33] the cloud and they communicate always [53:35] through the cloud. So that's a that's [53:36] the first thing. We don't right now [53:38] actively work on multiple drones you [53:40] know tracking the same the same subject [53:43] sort of in in that way. It is perhaps [53:45] something that, you know, when we are [53:46] going to see use cases we can think [53:47] about in the future. It's not something [53:49] that we are actively actively pursuing, [53:51] but most importantly it always goes [53:52] through the through the cloud. [53:54] They never communicate with each other, [53:55] yeah? Oops, sorry. [53:57] >> You have what with what models with [54:00] >> Yeah, definitely. [54:03] Yeah, definitely. I mean, obviously [54:06] everything that requires high frequency [54:08] runs on the drone itself, which is like [54:10] obstacle avoidance, for example. So, [54:12] that we generally that we run on the on [54:14] the vehicle itself. Uh subject tracking [54:16] we run on the vehicle as well. We I [54:18] showed that we have sort of a sky mask [54:20] as well to figure out where actually the [54:21] satellites are. All of those things are [54:24] on the vehicle itself. Generally, as you [54:26] as you said, I think that's a fair fair [54:28] guiding principle. Everything that [54:30] requires high frequency commands or like [54:33] sort of cannot has to work also if you [54:35] don't have connectivity. They are, you [54:37] know, even though we have very very good [54:39] connectivity, we still lose the lose the [54:41] connection time to time. So, it needs to [54:43] operate, so you can't crash into [54:44] building, you know, if if you lose the [54:45] connectivity. So, that all runs on the [54:47] vehicle. Uh but everything that is sort [54:49] of higher level reasoning or like like [54:52] open vocabulary detection type of [54:54] things, that can all run in the cloud. [54:57] Yeah. [54:57] >> Uh in the previous slide you were [54:58] talking about training articles on the [55:01] simulator. [55:01] >> Yeah. [55:02] >> So, [55:03] I wanted to first ask how did you sort [55:05] of build the simulator? Was it using [55:06] like the occupancy maps that you were [55:08] talking about? And did you face any [55:10] sim-to-real gap when you actually [55:12] deployed it? And if so, how did you [55:14] >> Yeah, that's awesome question. Uh so, [55:17] we do rely on on something called [55:19] digital surface surface map. So, it's [55:21] not necessarily like 3D 3D voxels, but [55:23] it's more like a height of every single [55:24] pixel. [55:26] And that is where we train all the [55:28] policy in simulation. Uh but when we [55:31] deploy it, we actually run we use the [55:33] same maps. So, there is technically very [55:35] little sim-to-real gap from that [55:37] perspective. Uh there is of course [55:39] sim-to-real gap, the fundamental [55:40] underlying the sort of vehicle dynamics [55:42] model that that we have seen, but I [55:43] think at this point of time we have [55:45] closed the gap fairly fairly well. So, [55:47] if you train a policy in your simulator, [55:49] you have very high chance that it's [55:51] going to work also also in real world. [55:53] The one thing I'll build on top of it, [55:55] obviously, like the one gap that we are [55:56] seeing right now, the the biggest gap we [55:58] are seeing right now is [56:00] is uh shorter occlusions or like [56:03] occlusions that perhaps are not in a [56:04] map. We build the maps, we reconstruct [56:06] the maps, so buildings are fairly well [56:08] accounted for, but something like trees, [56:11] and especially if you are far, you know, [56:12] far behind, that's actually is much more [56:14] difficult. So, that's the area where we [56:16] are seeing right now a biggest biggest [56:18] gap, the next thing we have to work on, [56:20] where we are going mainly in the [56:21] direction of more towards end-to-end and [56:22] those type of models where that problem [56:24] is going to be much more prominent. [56:26] Yeah. [56:27] >> Question, are there any [56:29] environmental [56:30] conditions where you cannot fly the [56:32] drone? Like, for example, there is a [56:34] heavy rain, [56:35] not possible. Fire in Los Angeles area, [56:39] not possible because of like [56:41] too much smoke, too much uh [56:43] >> Yeah. [56:44] So, uh the drones do have specs as any [56:47] single sort of aviation aviation um [56:50] uh [56:50] hardware. [56:51] So, if there is like thunderstorm and [56:54] extreme high winds, [56:55] we of course there's always a as much [56:57] trust that these systems can generate. [56:59] So, if there is very high wind, we [57:00] cannot actually move forward, then we [57:02] generally cannot fly. All of the docking [57:03] station have sort of thermal station, so [57:05] we know if these if there are harsh [57:07] environments, then we sort of prompt the [57:09] user that they shouldn't actually fly. [57:10] They can always override it, but we say [57:12] we shouldn't. So, that's one of the one [57:14] of the examples. The other one, for [57:16] example, in rain, we can still fly. All [57:18] of these devices are waterproof, so you [57:20] can fly in a rain, but uh we for example [57:23] turn off obstacle avoidance today [57:24] because what we have seen at the rain, [57:25] there is a lot of sort of spurious [57:27] voxels that we have to then uh that that [57:29] would result in in in movement of the [57:31] vehicle, so we turn off obstacle [57:32] avoidance in rain. But other than that, [57:34] I mean [57:35] uh you generally fly these beyond visual [57:37] line of sight, so you don't actually see [57:39] them, and of course you always uh [57:42] uh so from that perspective, there is [57:45] only like there are specs for the [57:46] vehicle, especially for for I would say [57:47] wind, the most important. [57:50] Yeah. [57:52] >> Yeah, you're extremely useful in the [57:53] cold. I was wondering how fast and how [57:56] slow [57:57] fly. [57:57] >> Yeah, it's around the the fastest one is [58:00] around 100 is 100 miles 100 miles per [58:02] hour. [58:03] Uh [58:04] we have uh if you anyone is interested, [58:06] there is we did a sprint in our Ascent [58:08] conference on YouTube. You can see [58:10] actually sprint that we were we reached [58:11] 100 miles. So it's 100 miles an hour. [58:13] The flight time [58:16] I would lie. I don't have the exact [58:17] number in my head. [58:19] Uh [58:20] around 60. I I don't remember the exact [58:23] number. [58:24] Uh but it's it's definitely much higher [58:25] than than [58:26] uh [58:27] than than than quadcopter. And how slow [58:31] uh [58:32] 20 [58:34] uh I would also lie to the exact number. [58:35] At some point, you know, of course you [58:36] have to de-transition. [58:38] Um so there is of course like the the [58:39] way it works that there is one one MPC [58:42] controller under the hood that handles [58:43] all of the transitions. And if you were [58:44] to command too slow speed, then it will [58:46] start de-transitioning and start moving [58:48] in horizontally. [58:50] Yeah. [58:51] >> Can you talk more about how like the [58:52] problem takes [58:55] Can you create a salt battery now? [58:57] >> Yes. [58:58] Yeah, [58:59] uh a good question as well. So for [59:01] defense use cases, we have a we have [59:03] somewhat different product, which is [59:05] called Xtend D, and that has different [59:07] it has a special different radio, most [59:09] importantly. It has different radio. It [59:11] doesn't have 5G. It doesn't have course [59:12] connectivity. So it has So that's the [59:15] that's the main area that that we are [59:16] seeing. And the other one is of course [59:18] also GPS denied. In these areas, they [59:21] generally jam GPS. But that I I do think [59:24] that something that is is to actually go [59:26] that it happen not only in sort of the [59:28] the military use cases that it can [59:30] actually happen in normal cities as we [59:31] are seeing for example now in in other [59:33] other areas. [59:35] Um but that's the other one. We have VIO [59:37] that would that generally works even in [59:39] very high altitude. So we we can still [59:41] rely on that and get [59:43] drifty more drifty pulse estimate [59:44] obviously as GPS, but it's generally [59:46] good enough to fly somewhere and return [59:48] in rough approximate where you are. [59:52] Good question. Yeah. [59:53] >> I have a question. Uh when it comes to [59:57] uh when it comes to the tracking um [60:00] some [60:02] um [60:03] How do you say that? So when it comes to [60:04] the tracking I'm I'm guessing that uh [60:06] the user is the one selecting the object [60:08] to track. [60:08] >> Yeah. [60:09] >> Um I know that in security when it comes [60:12] to drones used for security and uh some [60:14] of the customers might want the tracking [60:16] to be made more autonomously and for [60:18] example locking into uh an object or [60:21] especially someone whose behavior is [60:24] kind of suspicious. I know this is kind [60:26] of [60:27] um complicated because especially [60:29] because of the resolution because the [60:30] drone is flying very high. So I'm My [60:33] question is is it something you're [60:35] working on? Is it a project uh going on [60:37] and [60:38] >> Yeah. [60:38] >> What are the challenges that you [60:39] encounter? [60:40] >> Yeah, yeah, good question. So um there's [60:43] So the question was mainly like whether [60:45] we can detect uh detect movement and [60:47] suspicious activity of the individuals. [60:49] I mean that's definitely a very [60:51] interesting and very common use case [60:52] that we are hearing that we are also [60:53] actively actually working on. The part [60:55] that I think is is very important for us [60:57] is that and and what what is very [60:59] important for for everyone is that we [61:01] can control all parts of the stack. So [61:02] what we can do is that if we first [61:04] detect someone that we may be suspicious [61:06] or not, what we do even today we [61:08] actually zoom in, we punch in, we high [61:10] get higher crop or high resolution and [61:12] we confirm whether that's actually [61:14] personal or it was just false positive [61:16] and we so we get we and we get also like [61:18] a snippet like some some amount of [61:20] seconds of video and then we can confirm [61:22] whether that is something that we have [61:24] thought before. So, generally where we [61:26] can actually zoom in and zoom out, and [61:28] you know, if you go fast enough, you can [61:29] sort of see that the drone sort of zooms [61:30] in and zooms out to different areas and [61:32] the things that it feels that is [61:34] suspicious. [61:36] Uh yeah, yeah. [61:38] >> [snorts] [61:45] >> Yes. Yeah, I mean, that's a great [61:48] question. So, I mean, we work very [61:49] closely also with FAA uh make it [61:52] working, and we have we have a [61:53] regulatory in terms of generally helps [61:54] these customers to acquire permissions. [61:56] All of these customers and up their [61:58] pilots have sort of beyond visual line [62:00] of sight permissions, and at this point [62:01] of time they have even a permission to [62:03] operate multiple drones simultaneously. [62:06] And that's generally what we work with [62:07] these customers to acquire these acquire [62:10] these uh regulatory allowances. And they [62:12] still They are, of course, as I showed [62:14] you in the beginning, uh there are still [62:16] restrictions that you In some areas you [62:18] have to fly below certain altitude, and [62:20] that's generally all enforced by the by [62:22] the software as well. So, that's [62:23] generally quite nice thing that we can [62:25] actually enforce some of these [62:26] restrictions by the soft underlying [62:28] software that pilots cannot necessarily [62:30] override. And and and of course, the pay [62:32] pilot pilots are also trained and [62:33] certified to operate the part 107. Yeah. [62:37] >> How fast are the tracking drones? And [62:39] like, how much do you prioritize like [62:40] agility for [62:42] cars? [62:44] >> Uh so, [62:46] uh as an engineer, I don't know in miles [62:48] an hour. The drone [62:50] maximal the maximal safe speed for [62:51] example is 20 m per second. [62:53] Uh so, 20 m per second is generally the [62:54] maximal speed a drone can fly, and it [62:56] has obstacle avoidance. Uh we don't we [62:59] don't today necessarily change the [63:01] models for vehicle or cars. It's sort of [63:04] a unified thing. Uh so, we don't [63:06] necessarily differ or differentiate that [63:08] we we do very different very different [63:11] uh [63:11] behaviors for one or the other. It's [63:14] it's the same. [63:15] About 20 m per second, yeah. [63:17] >> I'm assuming that this is another thing [63:18] you're looking for a lot of time. [63:20] Yeah. [63:27] Yeah. [63:39] Uh yeah, good question. Uh [63:41] so, what we have seen actually people [63:43] also do to use the you know, they can [63:44] lower the altitude as well and they they [63:47] are you know, X10 is a modular device [63:49] that you can mount different [63:50] attachments. You can put there like we [63:52] have a spotlight, so you can actually [63:53] turn on the light and turn it off. You [63:55] have also speaker attachment to it, so [63:56] you can actually talk through it and [63:58] that's generally the way you can you [63:59] know, lower you can lower the altitude, [64:00] turn on the flashlight. These [64:02] flashlights are fairly strong, so they [64:03] are very very noticeable and and and [64:06] customers are using them quite actively [64:08] and that's generally the way. It [64:09] actually even for security use cases, we [64:11] have seen that uh that these drones can [64:15] be somewhat spooky. [64:16] So, if someone is an area that they [64:18] shouldn't be in, uh if they see a drone, [64:20] they generally like leave by themselves. [64:23] >> [cough] [64:23] [clears throat] [64:24] >> Yeah. [64:53] Yeah, uh [64:55] Right. [64:56] I mean, uh [65:00] As [snorts] I meant So, all of these [65:02] docking station So, um there is also way [65:04] to make these docking station remote [65:05] like more more a mobile. So, um [65:08] you can put these docking stations in a [65:10] in a trailer or whatever and you can [65:12] just move it anywhere where you want. [65:13] And they And they also come with [65:15] external radio. So, that's a way you [65:17] generally can go anywhere you want. You [65:18] park it. You sort of create a set up [65:21] flight, as we call to figure out the [65:22] locations, and you can rely on your [65:24] local local antenna, and fly purely [65:27] based on that. That's always possible. [65:29] Obviously, the you know, the the [65:31] distance that you can fly [65:33] is is is more limited in comparison if [65:35] you rely on 5G, but that's that's [65:38] something that you can always do. [65:49] Yeah. [65:54] >> [clears throat] [66:01] >> Yes. [66:10] Yeah. [66:22] Uh [66:23] it's good question. [66:25] I mean, it it will very much depend on [66:27] the use case you are trying to solve. I [66:30] do think we will have, as I mentioned, [66:31] for example, just like [66:33] if you transcribe down one call, [66:36] you just need to we just sort of open [66:37] the doc. That's, you know, can be one [66:39] part of the model or like doesn't need [66:40] to be necessarily VLA. [66:42] But definitely VLA [66:44] as a general solution is very [66:46] interesting for us, and it's definitely [66:47] an also an area that we are we are [66:49] actively working on. You know, it will [66:51] probably take us some time to [66:53] you know, mature it and getting all the [66:54] way to the fully functioning functioning [66:56] solution. [66:58] Um so, we'll do some intermediate [66:59] intermediate sort of more tailor-made [67:01] solutions such that we we enable some [67:03] parts of the functionality, but I I [67:05] think that some form of a larger [67:07] foundation model with sort of some [67:09] backbone that is pre-trained on large on [67:11] on internet-wide data is going to be is [67:14] very interesting for us. [67:20] Uh where do we store? I mean, we have a [67:22] fairly large substantial data for this [67:24] part of time that we could potentially [67:26] do like that. [67:31] Uh I think I think I mean, initially, [67:34] you know, given, you know, these [67:35] generally are multiple billions of [67:37] parameters models, I do think that [67:39] initially we will focus on the on the [67:41] cloud side and make sure that it it that [67:43] it works there first, but I do think, [67:45] you know, as a as a general technology [67:46] trend, we are all seeing that these [67:48] models are getting more efficient, [67:49] faster, and you know, over time you can [67:51] distill them perhaps to smaller, more [67:53] targeted targeted problem sets, but I do [67:55] think over time there is there is [67:57] definitely a path that these are going [67:58] to get on the drone. You know, it's not [68:00] going to be first application that we [68:01] will it's not going to be the first [68:02] thing that we will do, but I think over [68:04] time it's it's definitely going to [68:05] happen. [68:07] Um it very much it's it's roughly you [68:10] can think of it roughly 100 [68:11] milliseconds. [68:13] Yeah. [68:21] We do we do combination of all. Like [68:23] that's the that's the that's I guess the [68:25] the connect fusion plus that you know, [68:27] it's very important for us that we take [68:29] all of the signals and we sort of [68:30] smartly fuse all of the bandwidth from [68:33] all of the [68:34] all of the sort of signals together to [68:37] get the most information. [68:40] I mean, if if you drop one of them, you [68:42] can perhaps rely more on the other ones, [68:44] you know, we we do we do create [68:45] generally also maps for for customers [68:47] that you know, if we are doing any [68:48] larger deployment, we can sort of map [68:51] what's the cell coverage in the areas, [68:53] and then if they they know at least if [68:55] they fly somewhere, they would lose some [68:56] they would lose connectivity. Uh so, we [68:58] do that, but we generally like do just [69:00] smart smart we just fuse it smart ways [69:03] in smart ways. And yeah, some case it [69:05] happens if you know you for example lose [69:07] two and a half out of three or like [69:08] there is a drops [69:10] drop on all three of them. There is a [69:11] chance you will lose you will see cuts [69:13] in the video. That's also that also [69:15] happens. Because we fuse so many of [69:17] them, it happens less frequently but it [69:19] it it may happen still. [69:24] Cool. [69:25] Yeah, I think we are 10 minutes over. [69:28] But yeah, cool. Thank you.