## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=4s) Introduction and the Current Reality of Autonomous Drones
Sammy Hamilton, Vice President of AI at Heven Aerotech, introduces his work leading the AI and autonomy stack for hydrogen-powered unmanned aerial systems (UAS). The core objective of his team is to significantly remove human operators from the loop during vehicle operations. Hamilton addresses common misconceptions in the drone industry by contrasting current operational capabilities with popular science-fiction narratives.
Currently, basic drone functions—such as onboard perception, navigation, flight control, target detection, target classification, and automatic target recognition (ATR)—are fully solved engineering problems. Drones can execute flights and perform optical or computer vision sensing independently without maintaining a active connection to a remote operator. However, Hamilton stresses that the industry has not reached the popularized vision of autonomous swarms taking over the world. Modern autonomous systems operate by strictly following pre-programmed instructions—what Hamilton calls "executing a recipe"—without evaluating whether the plan is optimal or questioning its underlying purpose.
- **Speaker Role:** Sammy Hamilton, VP of AI at Heven Aerotech.
- **Platform Specialization:** Hydrogen-powered unmanned systems.
- **Proven Technologies:** Perception, GPS-independent navigation, flight control, detection, classification, and Automatic Target Recognition (ATR).
- **Current Operational Paradigm:** Drones strictly follow deterministic plans and recipes rather than generating independent strategies or altering operational goals mid-flight.
[00:218] The Autonomy Gradient: From Manual Control to Independent Judgment
Hamilton presents a framework called the **Autonomy Gradient** to classify the progression of drone capabilities from basic manual piloting to full machine independence:
1. **Manual Flight:** A direct 1:1 setup where a single human operator manually controls every movement of the aircraft using a physical remote control.
2. **Autopilot:** The aircraft follows predetermined spatial patterns—such as waypoints, ellipses, or figure-eights—set via a computer interface. The autopilot manages flight stabilization, collision avoidance, and power consumption monitoring to prevent crashes.
3. **Autonomy:** The aircraft senses its immediate environment and dynamically decides how to execute an assigned task. This level encompasses platform, mission, collaborative, and swarming autonomy.
4. **Independent Judgment:** The system interprets mission intent, establishes outcomes, dynamically determines execution steps, and reserves the authority to alter its objectives mid-flight.
Hamilton clarifies that the majority of the commercial and defense drone industry currently operates between levels 2 and 3. Level 4 remains unrealized; perception of its existence is largely fueled by promotional sizzle reels, marketing materials, synthetic media, and media coverage.
At a foundational level, drone autonomy mirrors human interaction through a **Perception $\rightarrow$ Cognition $\rightarrow$ Action** loop:
- **Perception (Sensing):** Uses computer vision, optical sensors, ATR, acoustic sensors (hearing), and vibration or thermal sensors (feeling).
- **Action (Actuation):** Requires precise coordination across physical components including elevators, servos, fuel cells, rotors, and electric motors.
- **Cognition (Decision-making):** In current fielded systems, cognition relies heavily on complex, hard-coded behavior trees—effectively structured "choose-your-own-adventure" logic paths with predefined outcomes. Newer approaches incorporate reinforcement learning and state-space models.
[06:08](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=368s) Physical and Environmental Constraints on Drone AI
Transitioning advanced AI models to physical drone platforms presents distinct engineering and environmental bottlenecks that do not exist in purely software-based domain applications. Hamilton highlights key technical constraints:
- **SWaP Limitations (Size, Weight, and Power):** Large compute platforms capable of running heavy modern AI models demand high power, generate significant heat, create mechanical vibration, and require dedicated cooling systems. Integrating bulky compute hardware onto small form-factor platforms—such as a 10-inch quadcopter weighing under 10 pounds—directly reduces payload capacity and flight duration.
- **Millisecond Control Loops:** Unlike autonomous ground vehicles (e.g., Waymo) that can safely pull over to the side of the road to re-evaluate sensor data or update maps, an airborne platform cannot pause mid-air. Sensor failure or logic stalls result in immediate platform loss due to gravity.
- **Sensing Difficulties:** Maintaining steady sensor output at high altitudes (e.g., 10,000 feet) on vibrating carbon-fiber airframes presents severe computer vision challenges compared to road-bound vehicles.
- **Denied and Disconnected Environments:** Drones frequently operate beyond visual line-of-sight (BVLOS), inside structures, or within GPS-denied and communications-denied environments where real-time cloud offloading or remote teleoperation is impossible.
- **Operator Ratios and Legal Constraints:** While the goal is a 1:Many operator-to-drone ratio, regulatory constraints (such as FAA safety rules) and risk-aversion often result in the inverse. Hamilton recalls observing a field demonstration where a single autonomous air vehicle required three human operators: one monitoring the command-and-control (C2) link, one watching the autopilot, and one monitoring the autonomy layer.
[10:10](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=610s) Emerging Technological Breakthroughs in Drone Autonomy
Hamilton highlights several practical innovations currently transitioning from research into fielded drone stacks:
- **Over-the-Air (OTA) Model Updates:** The capability to wirelessly push updated perception, ATR, and autonomy models directly to fielded drone fleets without physical tethering via Ethernet.
- **Translation of Commander's Intent:** Replacing complex, vendor-specific graphical interfaces (HMIs)—which require manual inputs for bounding boxes, no-fly zones, no-look zones, and ingress/egress routes—with Large Language Models (LLMs), natural language understanding, and agentic workflows. Operators can issue high-level instructions (e.g., *"Search this area of the coastline for the missing person"*), which agents automatically translate into executable mission parameters.
- **Counter-Counter-UAS (C-UAS):** Rapid software adaptation cycles to counter active kinetic and directed-energy threats, such as lasers, nets, shotguns, and electronic warfare.
- **Hardware Innovations:** The integration of low-power neuromorphic computing chips, dual-use directed energy systems, and edge-based Small Language Models (SLMs) optimized for local compute hardware.
[12:51](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=771s) Value Proposition: Human Augmentation and Risk Reduction
Integrating AI onto long-endurance platforms allows humans to offload tedious or hazardous tasks to machines. As an example, Heven Aerotech’s hydrogen-powered drones offer continuous flight endurance of 8 to 10 hours alongside low thermal and acoustic signatures. Human operators experience cognitive fatigue and loss of focus during repetitive tasks, such as staring at monitoring screens for hours or inspecting a multi-mile bridge span. Autonomous aerial platforms handle continuous monitoring without performance decay.
Crucially, autonomous systems reduce human physical exposure. Drones can be deployed into highly dangerous environments—such as burning or collapsed structures, earthquake rubble, active combat zones, or contested airspace—without placing human lives at risk.
[14:32](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=872s) Risk Allocation and Misplaced Authority
The adoption of AI-enabled autonomous drones introduces severe operational and ethical risks:
- **High-Consequence Machine-Speed Failures:** Because autonomous onboard loops operate at millisecond speeds, failures happen rapidly and are usually discovered post-crash. Hamilton draws an analogy to a public demonstration where a humanoid robot collapsed on stage and had to be dragged off by technicians; airborne platform failures carry far more destructive physical impacts.
- **Expanded Attack Surfaces:** Deploying onboard machine learning models opens platforms to model poisoning, visual ATR trickery, position spoofing, and GPS spoofing.
- **Misplaced Authority:** Hamilton identifies the delegation of operational authority as the primary risk facing autonomous deployment. The risk lies in delegating decisions to algorithms without fully evaluating the parameters or intent. As systems shift along the autonomy gradient toward independent decision-making, engineers and operators must explicitly define which responsibilities can be safely delegated to the software stack.
## Video description
Sammy Hamilton - VP of AI, Heven Aerotech
AI isn't a feature you bolt onto a drone. It's the decision that determines whether your platform can operate independently, even when connectivity is lost and there's no one at the controls. In this session, I'll walk through how AI is reshaping drone technology from the ground up: onboard perception, navigation without GPS, real-time decision-making, and multi-vehicle coordination. I'll address the critical question head-on: should drones be AI-enabled, and to what extent? I'll draw from hands-on experience building AI and autonomy capabilities for hydrogen-powered UAS platforms serving the global market. I'll cover what AI can do today for autonomous drones, what it can't, and why that distinction matters. You'll leave with a clear picture of what's real, what's next, and what's still missing in what may be AI's fastest-moving application domain.
可能性的艺术：Heven Aerotech 与 AI 赋能无人系统
Heven Aerotech 人工智能副总裁 Sammy Hamilton 指出，尽管机载感知与目标识别已基本成熟，但现有无人机大多仍处于执行预设程序的阶段，介于自动驾驶与自主控制之间，尚未达到宣传中的独立判断水平。在无人机上部署 AI 面临尺寸、重量与功耗（SWaP）限制、毫秒级控制延迟及通信受限等物理挑战；尽管前沿大语言模型智能体正推动以自然语言简化航线规划，将操作权限委托给算法仍伴随着即时坠毁风险以及模型投毒、对抗性攻击等安全挑战。
Heven Aerotech 人工智能副总裁 Sammy Hamilton 指出，目前机载感知、无 GPS 导航和自动目标识别（ATR）已基本成熟，但现有无人机仍严格遵循确定性指令运行，行业尚处于执行预设程序的阶段，尚未达到能独立调整任务目标的水平。 (0:04)
Hamilton 提出了包含手动控制、自动驾驶、自主控制和独立判断四个阶段的自主梯度框架；目前大多数商用和军用无人机仅介于自动驾驶与自主控制之间，最高阶的完全独立判断主要仍存在于宣传物料中。 (3:38)
无人机部署 AI 面临严格的物理约束，包括尺寸、重量和功耗（SWaP）限制、毫秒级控制回路对计算停顿的零容忍，以及在无 GPS 和通信受阻环境下的离线计算挑战。 (6:08)
前沿技术正推动 OTA 模型无线更新，并通过大语言模型（LLM）与智能体工作流将指挥官的自然语言意图直接转化为航线规划，逐步取代传统繁琐的人机交互界面。 (10:10)
将操作权限过渡委托给算法是无人机自主化的主要风险所在，机载毫秒级决策一旦发生故障将造成即时坠毁，同时还面临对抗性攻击、模型投毒和欺骗等多重安全挑战。 (14:32)
Reading summary
Sammy Hamilton, Vice President of AI at Heven Aerotech, argues that while perception and flight control are solved engineering problems, operational drone autonomy remains largely confined to deterministic rules rather than genuine independent machine judgment.
The Autonomy Gradient and Current Industry Reality
Hamilton outlines an Autonomy Gradient spanning four operational stages: manual flight, autopilot, autonomy, and independent judgment. In current commercial and defense deployments, systems generally function between level 2 (autopilot) and level 3 (autonomy). Level 4—where an unmanned aerial system (UAS) independently interprets mission intent, establishes outcomes, and dynamically alters objectives mid-flight—remains unrealized, despite promotional materials and synthetic media suggesting otherwise.
Onboard operations mirror a perception-cognition-action loop. While perception via optical sensors, computer vision, and automatic target recognition (ATR) is well-established, Hamilton explains that cognition in deployed systems still relies primarily on hard-coded behavior trees. Rather than formulating novel strategies, drones execute deterministic recipes without evaluating the broader purpose or optimality of their assigned plans.
Physical and Environmental Constraints on Aerial AI (6:08)
Transitioning artificial intelligence onto airborne platforms introduces physical bottlenecks absent in pure software or ground-based autonomous vehicles. Size, weight, and power (SWaP) constraints restrict the hardware that small airframes can carry; heavy compute modules generate heat and vibration while reducing payload capacity and flight times. Furthermore, drones cannot pause mid-flight like ground vehicles to recalculate sensor feeds—millisecond control loop failures result in immediate platform crashes.
Operational environments further complicate deployment. Drones often operate beyond visual line-of-sight (BVLOS) in communications-denied or GPS-denied spaces, preventing real-time cloud offloading. Regulatory requirements also hinder efficiency; Hamilton notes that instead of achieving a 1:Many operator-to-drone ratio, risk-averse setups sometimes invert the ratio, requiring three human monitors for a single autonomous vehicle.
Emerging Onboard Workflows and Practical Value (10:10)
Technological advancements are moving from research into fielded drone architectures. Key developments include over-the-air (OTA) model updates for deployed fleets, edge-based Small Language Models (SLMs), neuromorphic computing chips, and rapid software countermeasures against kinetic or electronic counter-UAS systems. Additionally, agentic workflows and Large Language Models are beginning to replace complex graphical human-machine interfaces, allowing operators to issue high-level mission intent in plain language that software translates into flight parameters.
According to Hamilton, these systems provide concrete value by offloading hazardous and cognitively exhausting tasks from human operators. Hydrogen-powered platforms can sustain 8 to 10 hours of flight without performance degradation, conducting extended monitoring or penetrating dangerous areas like collapsed buildings, active fires, and contested combat zones without risking human life.
Operational Risks and the Threat of Misplaced Authority (14:32)
Deploying autonomous systems introduces severe failure modes, primarily because machine-speed errors occur within milliseconds and are often identified only after a crash. Edge-deployed machine learning models also expand the platform's attack surface, leaving systems vulnerable to model poisoning, visual trickery, and spoofing.
Hamilton emphasizes that the most critical challenge is misplaced authority. As systems progress along the autonomy gradient, the primary operational risk lies in delegating decision-making power to algorithms without adequately defining parameters, understanding algorithmic limits, or maintaining appropriate supervisory boundaries.
While the drone industry continues to expand endurance and edge intelligence, the core unresolved challenge remains establishing strict boundaries around delegated authority before platforms are granted greater operational independence.
## Background
本视频由频道 Ai4 于 2026 年 8 月 18 日发布在 YouTube 平台。内容为 Heven Aerotech 的 AI 副总裁 Sammy Hamilton 带来的主题演讲，围绕 AI 如何从底层重构无人机系统展开，重点剖析机载感知、无 GPS 导航、实时决策与多机协同，深入探讨无人系统引入 AI 的合理边界与现实落差。
## TL;DR
演讲系统阐述了无人系统从手动飞行到独立判断的“自主性梯度”，指出当前技术正处于从预设自动驾驶向任务自主跨越的阶段，远未达到完全独立思考的科幻水平。Sammy 指出无人机受限于 SWaP（尺寸、重量与功耗）、毫秒级控制回路以及高空无法“靠边停车”等物理与通信瓶颈，决策核心目前仍以硬编码行为树为主。尽管空中无线模型更新、自然语言指令翻译及神经拟态芯片等前沿技术不断突破，但在国防等高风险场景下，必须警惕 AI 攻击面扩大与决策权错位的风险。他最终提出明确的人机分工底线：由回路中的人类决定战术意图与最终结果，将具体的飞行与感知执行交由机器完成。
## Atmosphere & style
整场演讲风格严谨务实、直击要害，主讲人以一线技术专家的冷静视角撕下了市场营销与科幻滤镜，客观厘清了技术现实与宣传误区。叙事从严密的底层工程约束层层推进至高维度的伦理与决策授权，节奏紧凑且充满思辨张力。结尾的观众问答直面俄乌、中东战场冲突带来的激进创新与本土审慎信任之间的矛盾，给整场分享增添了浓厚的地缘现实感与深沉的战略反思。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=4s) 现阶段无人机 AI 的基础感知、导航与自动目标识别（ATR）已高度成熟，但整体自主性仍停留在执行预设指令的“配方”阶段，远未达到科幻级别的自主思考与自我迭代。
- [02:18](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=138s) 自主性梯度模型将无人系统分为手动飞行、自动驾驶、任务自主和独立判断四个阶段，目前行业主力技术处于第 2 到第 3 阶段之间，真正的“独立判断”仍存在巨大差距。
- [04:28](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=268s) 无人机自主系统遵循“感知-认知-行动”闭环，当前可靠落地系统的认知决策核心本质上仍依赖复杂的硬编码行为树，而非完全自由的自我演化模型。
- [06:08](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=368s) SWaP（尺寸、重量与功耗）限制和重力物理法则构成了严苛瓶颈：无人机在空中发生毫秒级控制失效时无法“靠边停车”，必须在极低容错率下运行。
- [06:08](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=368s) 在超视距（BVLOS）、无 GPS 及通信中断等极端环境下，无人机无法像自动驾驶汽车那样寻求远程人工协助，必须具备完全独立的机载上下文理解与决策能力。
- [06:08](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=368s) 受制于空域安全监管与坠机风险担忧，当前部分自主无人机甚至出现“单机配备多名监控人员”的现象，人力绑定机制尚未真正解绑。
- [10:10](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=610s) 空中无线模型更新（OTA）与基于大语言模型的“指挥官意图翻译”等前沿技术，正在大幅降低操作门槛并加快野外 AI 模型的迭代速度。
- [12:51](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=771s) 自主化技术的核心价值在于释放人类认知负荷，氢动力长航时无人机能够替代人类在高疲劳（如数小时巡检）和高危险（如震后搜救、高危空域）场景下作业。
- [14:36](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=876s) 过度授权伴随“机器速度”下的不可逆失效、对抗样本与数据毒化等 AI 攻击面扩大的安全风险，最大的威胁在于将人类尚未厘清的决策盲目托付给算法。
- [16:41](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1001s) 在人命关天的防御场景中，人机权责分工必须守住核心底线：坚持由“回路中的人”设定战术意图与最终结果，机器仅负责具体的动作与路径执行。
- [17:33](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1053s) 现代冲突（如乌俄、以朗战场）推动了前线激进的技术创新，而平时国防体系更偏向审慎验证算法信任度，地缘对抗压力随时可能打破这一平衡。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=4s) 现阶段无人机 AI 的现实与误区
Heven Aerotech 的 AI 副总裁 Sammy Hamilton 介绍了团队的核心任务：为氢动力无人系统打造 AI 与自主系统堆栈（autonomy stack），力求大幅减少人类在操作环路中的参与（human-in-the-loop）。
讨论首先明确了无人机技术的现状与常见误区。在当下的技术水平下，感知（perception）、导航（navigation）和飞行控制（flight control）等基础功能已经相当成熟且高度可靠，无人机或航空器完全可以在不连接地面操作员或遥控器的状态下独立飞行。此外，计算机视觉与传感器技术的突飞猛进，使得目标自动识别（ATR, Automatic Target Recognition）、检测与分类等功能在现实中得到了高度应用。
然而，很多人会从“自主性”直接联想到概念图里那些科幻感十足、能够完全自主思考并掌控世界的无人机蜂群，这与现实相去甚远。目前的自主系统本质上只是在“执行配方（recipe）”——赋予它一套既定指令，它就会尽最大努力去执行，既不会擅自改进配方，也不会质疑这是否是最优解，更不会去思考执行任务的目的。
- **感知与导航**：基础的飞行控制与环境感知（如自动目标识别 ATR）已被完全攻克并广泛落地。
- **任务执行机制**：目前的自主化仅停留在按照预设逻辑高效执行规划，缺乏灵活变通与自我思考的能力。
---
### [02:18](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=138s) 自主性梯度：从手动遥控到独立判断
为了清晰阐述自主化的发展历程，Sammy Hamilton 提出了“自主性梯度（Autonomy Gradient）”模型，将无人系统的自主化程度划分为四个演进阶段：
```mermaid
accTitle: 无人系统自主性梯度模型
flowchart LR
subgraph S1["传统控制"]
A["1. 手动飞行<br/>(单人单机遥控)"] --> B["2. 自动驾驶<br/>(航点/预设路线)"]
end
subgraph S2["先进自主"]
B --> C["3. 任务自主<br/>(感知环境/蜂群协作)"] --> D["4. 独立判断<br/>(自主决定意图与决策)"]
end
```
1. **手动飞行（Manual Flight）**：传统的“一机一遥控器”模式，所有飞行与决策完全依赖人类操作员。
2. **自动驾驶（Autopilot）**：通过计算机界面预设航点（Waypoints）、椭圆路线或八字飞行轨迹。系统能够进行针对性的飞行调校，并具备基本的安全稳定保障（如防止坠毁、避障或电量/燃料保护），但依然只是在重复人类既定的指令。
3. **自主控制（Autonomy）**：无人机能够实时感知周边环境，并自主选择最佳方式来完成预设任务。这一阶段涵盖了平台自主、任务自主、协作自主（collaborative autonomy）以及蜂群技术（swarming）。
4. **独立判断（Independent Judgment）**：无人机能够自行理解任务的本质与意图，自行决定预期结果和实现路径，甚至能在飞行途中随时改变主意与行动方案。
目前绝大多数行业实践都处于第 2 到第 3 阶段之间，许多创新型企业正在大力拓展协作自主与蜂群技术的边界。然而，由于博眼球的宣传片、过度的营销手段以及失真的 AI 生成图片和新闻宣传，许多行业外人士误以为技术已经达到了第 4 阶段，但实际上距离真正的独立判断还有巨大的差距。
---
### [04:28](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=268s) 感知-认知-行动循环及其实现机制
无人机自主系统的运行逻辑与人类以及其他智能体非常相似，均遵循“感知（Perception）- 认知（Cognition）- 行动（Action）”三元闭环：
- **感知（Sensing）**：通过光学传感器与计算机视觉进行“视觉”感知，利用声学传感器进行“听觉”感知，并结合振动与热成像传感器实现对环境状态的全面捕获。
- **认知（Cognition）**：整合传感器数据，首先解决最核心的生存问题——“如何不从空中掉下来”，其次计算如何飞往下一个航点或完成路线规划，同时评估天气等外部环境对任务的影响。
- **行动（Action）**：相比人形机器人或自动驾驶汽车，小型无人机的执行机构更为复杂，需要升降舵、舵机、燃料电池、旋翼与电机等多组件的高精度协同工作，才能确保飞行安全。
在目前已部署且可信的自主解决方案中，“认知”层面的决策核心实际上树立在极其复杂的“硬编码行为树（behavior tree）”之上。虽然强化学习（RL）和状态模型（state models）等新技术正在迅速发展，但现阶段的决策机制本质上仍像一本“由预设逻辑构成的‘选择你自己的冒险’图书”，其演进和终止的方式早已被严格限定。
---
### [06:08](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=368s) 限制自主性发展的技术与物理瓶颈
制约无人机从预设逻辑向高级人工智能演进的原因，包含了严苛的技术与物理约束：
- **SWaP 限制（尺寸、重量与功耗）**：运行先进 AI 模型需要庞大的计算资源，这些硬件不仅体积大、耗电高，还会产生巨大的噪音与散热需求。若要将此类算力强行塞入一台重 10 磅、10 英寸的小型四轴飞行器中，势必会严重牺牲载荷与航程。
- **毫秒级控制回路与无法“靠边停车”**：自动驾驶汽车在遭遇系统故障或路线不明时可以随时靠边停车、查阅手册或重新定位；但无人机无法在空中暂停。物理定律（尤其是重力）是极其残酷的，一旦控制回路出现延迟或失败，轻则损失昂贵设备，重则从数百英尺高空坠落引发严重安全事故。
- **感知与通信难题（无法“打电话求助”）**：无人机常运行在远距离无视线（BVLOS）、无 GPS 或信号屏蔽等恶劣环境中（如巡检地下隧道、建筑物内部或关键基础设施）。自动驾驶汽车在遇到复杂路况时可以通过呼叫中心寻求远程人工协助，但无人机在毫秒级决策要求及通信中断的场景下无法“打电话求助”，必须具备高度独立的上下文理解能力。
- **“一比一”甚至“一比多”的人力绑定配比**：目前普遍采用“一个操作员控制一台无人机”的模式。Sammy 提到近期参观某顶级自主技术公司的演示时，现场甚至出现了“一台自主飞行器配置三名操作员”的情况（一人监控 C2 数据链，一人监控自动驾驶仪防止自主系统失效，一人监控自主系统本身）。这种现象主要源于联邦航空管理局（FAA）严格的法规安全限制，以及企业对坠机风险和空域许可权丧失的担忧，反而造成了人力的退化。
---
### [10:10](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=610s) 领域前沿：从超视距更新到自然语言意图传达
尽管前路充满挑战，近年落地的新技术依然为无人机领域带来了重大突破：
- **空中模型无线更新（Over-the-Air Model Updates）**：支持将最新的 AI 模型、ATR 模型、感知模型及自主模块一键部署或重新部署至已在野外作业的无人机编队中，无需将飞行器召回并插上网线手动调校，使硬件上的 AI 迭代速度能够媲美云端软件。
- **指挥官意图翻译（Translated Commander's Intent）**：传统的自主系统要求操作员在复杂的 HMI 界面上逐个手动点击设置禁飞区、观察区、进出航路等参数，不仅操作繁琐，而且掌握技术的人往往不是任务发起者。如今结合大语言模型（LLM）、自然语言理解与 Agent 智能体工作流，指挥官只需下达“搜索某海岸线区域寻找失踪人员”的口头指令，系统就能自动将其转化为具体的自主控制命令。
- **反制与反反制博弈（C-UAS Cat and Mouse Game）**：针对无人机的激光、防空网、猎枪等反制手段层出不穷，倒逼 AI 驱动的自主避障与反反制技术快速进化，成为行业中演进最快的细分领域之一。
- **新兴硬件与边缘计算**：神经拟态芯片（Neuromorphic chips）大幅降低了运行高级模型的门槛；边缘侧小语言模型（Edge-based SLMs）的轻量化也让小型无人机承载复杂的智能推理成为可能。
---
### [12:51](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=771s) 释放人类认知负荷与高风险暴露
人工智能与自主化技术的核心价值，在于人机分工——让机器处理重复、高耗能且危险的工作，从而将人类从繁重的认知负荷与安全风险中解放出来。
以 Heven Aerotech 开发的氢动力无人机为例，其具备长达 8 至 10 小时的超长续航能力，且热源与噪音极小。在进行长达数小时的长航时巡航或长达三英里的桥梁结构巡检时，人类操作员不可避免地会在 45 分钟后产生疲劳、走神、饥饿或注意力下降，无法保证从头至尾保持相同的检查标准；而自主系统则能始终如一地执行高精度巡视。
更重要的是**降低危险暴露风险**。在地震后的建筑物塌方废墟、火灾现场、高辐射区域以及高风险的对峙空域中，具备可信自主能力的无人机可以替代搜救人员或士兵深入险境，从根本上保障人员生命安全。
---
### [14:36](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=876s) 决策授权的潜在风险与权衡
在赋予无人机更高自主权的同时，必须深刻审视其带来的多重风险：
- **后果不可逆与“机器速度”下的失效**：当自主系统在毫秒级的闭环中以“机器速度”做出错误决策时，人类往往无法阻止过程，只能在事故发生后面对既成事实（例如此前人形机器人在舞台上瞬间瘫痪倒地，需数人将其抬下场；而高速飞行的无人机发生故障时往往只会以坠毁收场）。
- **AI 攻击面扩大（Attack Surface）**：每一项新引入的 AI 能力都构成了新的被攻击点，包括数据模型毒化（model poisoning）、位置与 GPS 欺骗、以及针对 ATR 模型的视觉对抗攻击（visual trickery）。平台的自主化程度越高，被恶意操控的潜在风险就越大。
- **错位的授权（Misplaced Authority）**：真正的风险不在于让机器做决定，而在于“让机器做出了人类尚未彻底想清楚的决定”。随着技术沿“自主性梯度”向独立思考演进，人类正在将决策权逐步让渡给机器。在此过程中，必须极其审慎地明确任务的“真实意图”、“预期结果”与“实现方式”，并精准定义其中哪些环节可以安全地委托给机器执行。
### [16:41](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1001s) 自主权梯度的抉择：让“人”定意图，让“机器”做执行
在无人机技术的十字路口，当下的每一个抉择都将深远影响该领域未来 5 到 10 年的演进轨迹。Sammy Hamilton 强调，如果行业犯下灾难性的错误，或者推进速度盲目超越了“受信任技术”（trusted technology）的发展节奏，不仅无法驱动突破，反而可能引发毁灭性的倒退。面对两个看似毋庸置疑的问题——“无人机应该具备 AI 能力吗？”以及“无人机应当实现自主化吗？”答案固然是显而易见的“当然应该”。然而，真正复杂且棘手的问题在于：授权的尺度到底在哪里？在这个从完全受控到完全自主的“权威与自主权梯度”（authority or autonomy gradient）上，人类究竟应该向机器下放多少决策权力？
作为常年在国防、军事及国土安全等全球最高风险（highest stakes）场景下研发与应用无人机系统的一线主导者，Sammy Hamilton 直截了当地指出，当前的 AI 技术远未达到可以把生死攸关的行动决策与航线策略全权托付的阶段。在人命关天的防御场景中，容错率几乎为零。他给出了一个极其明确的划分基准：“让回路中的人（human in the loop）来决定意图和最终结果，让机器人去负责具体执行”（"let the human in the loop determine the intent and the outcome and let the robot handle the execution"）。这一分工既充分利用了机器在路径感知与飞行控制上的执行力，又牢牢守住了道德与安全的核心边界。
```mermaid
flowchart LR
accTitle: 权威梯度与人机分工模型
subgraph HumanAuthority ["人类回路 (Human-in-the-Loop)"]
Intent["设定战术意图 (Intent)"]
Outcome["决定最终结果 (Outcome)"]
end
subgraph MachineAutonomy ["机器自主权 (Robotic Execution)"]
Perception["环境实时感知"]
Pathing["无GPS导航与避障"]
Execution["战术动作执行 (Execution)"]
end
HumanAuthority -->|"授权边界 / 策略指令"| MachineAutonomy
```
- **行业窗口期**：未来 5 到 10 年内做出的架构和信任规范决定了无人系统整个生态的技术基调。
- **自主权底线**：反对在现阶段将关键决策链和行动方针（course of action）全权委托给算法模型。
---
### [17:33](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1053s) 问答环节：战时激进创新与和平时期的“信任迟疑”
在演讲尾声的提问环节中，现场观众抛出了一个颇具哲理性且切中要害的问题：鉴于当今技术挑战的本质大多是人类政治与地缘冲突问题，在复杂的地缘格局下，我们把这件事“做对”（规范、稳妥而非鲁莽推进）的概率到底有多大？针对这一关于是否会“轻举妄动”的质询，Sammy 结合当天早些时候的研讨话题，深入探讨了“战时”与“平时”在技术落地逻辑上的根本分野。
Sammy 分析道，当前世界局势动荡，在乌克兰与俄罗斯、以色列与伊朗等深陷战火的国家和地区，战场的极度残酷性迫使前线爆发了极快、极其激进的无人机与 AI 创新应用。在这些生死存亡的实战中，迫切的需求压倒了对长远隐患的顾虑。然而，当这些在战火中催生的前沿技术被引入美国本土时，美国的国防工业基础（Defense Industrial Base）往往表现得相对谨慎和克制。业界会不断拷问：“我们能够信任这项技术吗？”“我们真的有必要在此刻就全盘采用吗？”“我们是否有充裕的时间去通盘评估其背后的潜在连带影响？”
在 Sammy 看来，美国当前的审慎是在“刻意放慢脚步，争取时间”以打磨可靠、可解释且值得信赖的算法；但现实的残酷性在于，外界风险的激增以及对快速应对能力的迫切需求，随时都可能打破这种审慎平衡，逼迫人们在受信任算法完全成熟之前就被迫承担未知的技术风险。演讲在主持人的致谢与全场掌声中圆满收尾。
- **战时驱动 vs. 平时审慎**：战区（乌俄、以朗战场）催生出高速迭代的实战创新，而美国国防工业体系则更侧重算法的信任度验证、合规审查与战略审度。
- **核心博弈**：安全审查所争取到的“时间”，随时可能被外界地缘风险激增所带来的“技术跟进紧迫性”所打破。
## Notable quotes
> AI isn't a feature you bolt onto a drone. It's the decision that determines whether your platform can operate independently, even when connectivity is lost and there's no one at the controls. — Sammy Hamilton
> [00:04](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=4s) 赋予它一套既定指令，它就会尽最大努力去执行，既不会擅自改进配方，也不会质疑这是否是最优解，更不会去思考执行任务的目的。 — Sammy Hamilton
> [06:08](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=368s) 一台自主飞行器配置三名操作员 — Sammy Hamilton
> [10:10](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=610s) 搜索某海岸线区域寻找失踪人员 — Sammy Hamilton
> [12:51](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=771s) 让机器处理重复、高耗能且危险的工作，从而将人类从繁重的认知负荷与安全风险中解放出来。 — Sammy Hamilton
> [14:36](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=876s) 真正的风险不在于让机器做决定，而在于‘让机器做出了人类尚未彻底想清楚的决定’。 — Sammy Hamilton
> [16:41](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1001s) let the human in the loop determine the intent and the outcome and let the robot handle the execution — Sammy Hamilton
> [16:41](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1001s) 让回路中的人来决定意图和最终结果，让机器人去负责具体执行 — Sammy Hamilton
> [17:33](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1053s) 我们能够信任这项技术吗？ — Sammy Hamilton
> [17:33](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1053s) 我们真的有必要在此刻就全盘采用吗？ — Sammy Hamilton
> [17:33](https://www.youtube.com/watch?v=JmgRrTBGNdw&t=1053s) 我们是否有充裕的时间去通盘评估其背后的潜在连带影响？ — Sammy Hamilton
## Mentioned
Sammy Hamilton, Heven Aerotech, Ai4, FAA, LLM, SLM, ATR, BVLOS, C-UAS, SWaP, 乌克兰, 俄罗斯, 以色列, 伊朗, 美国国防工业基础 (Defense Industrial Base)
## Video description
Sammy Hamilton - VP of AI, Heven Aerotech
AI isn't a feature you bolt onto a drone. It's the decision that determines whether your platform can operate independently, even when connectivity is lost and there's no one at the controls. In this session, I'll walk through how AI is reshaping drone technology from the ground up: onboard perception, navigation without GPS, real-time decision-making, and multi-vehicle coordination. I'll address the critical question head-on: should drones be AI-enabled, and to what extent? I'll draw from hands-on experience building AI and autonomy capabilities for hydrogen-powered UAS platforms serving the global market. I'll cover what AI can do today for autonomous drones, what it can't, and why that distinction matters. You'll leave with a clear picture of what's real, what's next, and what's still missing in what may be AI's fastest-moving application domain.
[Source](https://www.youtube.com/watch?v=JmgRrTBGNdw)
Source quotes
Sammy Hamilton: AI isn't a feature you bolt onto a drone. It's the decision that determines whether your platform can operate independently, even when connectivity is lost and there's no one at the controls.
[0:04] Sammy Hamilton: 赋予它一套既定指令，它就会尽最大努力去执行，既不会擅自改进配方，也不会质疑这是否是最优解，更不会去思考执行任务的目的。
[6:08] Sammy Hamilton: 一台自主飞行器配置三名操作员
[10:10] Sammy Hamilton: 搜索某海岸线区域寻找失踪人员
[12:51] Sammy Hamilton: 让机器处理重复、高耗能且危险的工作，从而将人类从繁重的认知负荷与安全风险中解放出来。
[14:36] Sammy Hamilton: 真正的风险不在于让机器做决定，而在于‘让机器做出了人类尚未彻底想清楚的决定’。
[16:41] Sammy Hamilton: let the human in the loop determine the intent and the outcome and let the robot handle the execution
[16:41] Sammy Hamilton: 让回路中的人来决定意图和最终结果，让机器人去负责具体执行
[17:33] Sammy Hamilton: 我们能够信任这项技术吗？
[17:33] Sammy Hamilton: 我们真的有必要在此刻就全盘采用吗？
[17:33] Sammy Hamilton: 我们是否有充裕的时间去通盘评估其背后的潜在连带影响？
Original transcript
[0:04] Um, so my name is Sammy Hamilton. I'm the VP
of AI at Heaven Aerotch. And at Heaven Aerotch,
[0:08] we build hydrogen-powered unmanned systems.
[0:11] Um, and my team owns the AI and autonomy stack. So
we're working on removing the human significantly
[0:17] in the loop for these autonomous systems. So
let's talk about drones. Next slide, please.
[0:23] Um, so we're going to start with where
we are today, what it actually is,
[0:28] and more specifically what it is not and where we
are not. And then we're going to talk about what's
[0:33] coming. Some of the things that are super exciting
to me as of being a practitioner in this industry.
[0:37] And then we'll talk about what we gain when we
get it right. But even more importantly, what we
[0:42] risk when we get it wrong. And I'll close out with
one question I'd like everyone to think about as
[0:47] they leave here. Thinking about the future of
these autonomous systems. Next slide, please.
[0:54] So where we are and where we are not. The first
thing to understand is that the basics work and
[0:59] they work well. Perception, navigation, flight
control, these are solved problems and a drone,
[1:06] really any aircraft, can fly without needing to
be connected to an operator or a person with a
[1:11] remote control. These things work well. Detection
classification, they also work well. The sensor
[1:17] industry, the computer vision industry has come
so far in the past several years that these things
[1:23] are real. Automatic target recognition real. It
works very well. Um, what's not real is the leap
[1:28] that people go from autonomy to the lovely AI
generated picture I have on the slide here of
[1:34] futuristic looking swarms of intelligent drones
doing whatever it is they want to do and taking
[1:40] over the world, right? Um, so the way autonomy
works today is it executes a plan. You give it
[1:45] a recipe. It's not going to juzj up the recipe.
It's not going to change it. It's not going to
[1:50] ask if it's the most optimal recipe. And it's
not going to ask why you're cooking. It's going
[1:54] to execute the recipe as best as it can. And that
is that. Um, and we need to have the confidence in
[2:00] this autonomy that these vehicles aren't going to
fall out of the sky onto our heads. Um, so where
[2:05] exactly are we on the spectrum between following
a recipe and our cool uh AI enabled, really smart,
[2:12] intelligent decision-making drone system in
the fictional picture. U next slide, please.
[2:18] So I call this the autonomy gradient. It's how I
organize my thoughts on where we started, where
[2:23] we are, and where we're headed. Um, so we'll start
on the left and move left to right. First, manual
[2:29] flight. So this is what we all think about uh
remote control operation. one operator, one drone,
[2:34] one remote control and that person is flying the
aircraft. They are making all the decisions. They
[2:40] are doing the flying. Then we get to autopilot. So
the way an autopilot generally works is you have
[2:45] a computer interface and you mark out behaviors
you want to see. You have way points. You have
[2:50] ellipses. You can do figure eights. But these are
all very predetermined patterns on an autopilot
[2:56] that's highly tuned to the aircraft itself.
There are some safety things and stabilization,
[3:02] right? It's going to do its best not to fall out
of the sky. It's going to do its best not to crash
[3:06] into other things. Um, it's going to do its best
not to run out of battery power or fuel. Um, but
[3:12] it is still just doing what we've already told it
to do. Um, then we get to autonomy. And this is
[3:17] where it gets a little bit more interesting.
Um, this is where the aircraft perceives its
[3:22] environment and then chooses what it's going to do
to execute the task at hand. And I'm going to lump
[3:27] platform autonomy, mission autonomy, collaborative
autonomy, and swarming in this autonomy category,
[3:33] understanding that some of those things are
much easier than others. Um, and then there's
[3:37] the leap to the independent judgment. This is
where the platform decides what the mission
[3:41] means. It decides intent. It decides
outcome. It decides how it's going to
[3:46] do all of these things. And it reserves the right
to change its mind and change its course of action
[3:50] in the middle of a flight. And this is 100%
absolutely where we are not today with autonomy.
[3:56] Um so the point of this is that most of the
industry is operating somewhere between two and
[4:00] three. Um we have a lot of really innovative
companies out there pushing the boundary on
[4:04] collaborative autonomy and on swarming. Um but
I do think that level four is where a lot of
[4:10] non-industry practitioners believe we are based
off cool sizzle reels, really great marketing,
[4:16] misleading AI images and and news reels, right? we
all think we're we also have of course propaganda
[4:21] as well. Um so we're not close. Um but let's
take a step back and talk about how autonomy
[4:28] actually works. Um it works a lot like humans do.
Perception, cognition, action. So I'm up here. I'm
[4:34] sensing my environment. I'm going to think about
what I would like to do and then I'm hopefully
[4:39] right going to do it. Uh and this makes really
intuitive sense when we think about humanoid
[4:43] robots or self-driving cars because we all
understand fundamentally how those things work.
[4:48] But what does this look like when we're on
a small drone? We still do the same things,
[4:52] right? We sense the environment. We see with
computer vision, optical sensors, ATR. We hear
[4:57] with acoustic sensors. We feel with vibration
sensors and with thermal sensors. And then we
[5:02] piece it all together in its most basic form.
What do I need to do not to fall out of the air?
[5:07] That's most important thing. Um, and then what do
I need to do to accomplish what I set out to do?
[5:11] How am I going to get to the next waypoint? Once
I get to that way point, how am I going to travel
[5:17] in the ellipse? Is there weather or other things
that are impacting my ability to accomplish this?
[5:22] And then we get to act. And in my mind, this is
more complicated again because we all understand
[5:26] how humanoids work. We most of us understand how
cars work. In these small drone form factors,
[5:31] you have elevators and servos and fuel cells
and rotors and motors and many things that need
[5:36] to work collaboratively collaboratively together
to again make sure the aircraft doesn't fall out
[5:42] of the sky. Let's double click on the cognition
piece. The thinking right now with most fielded
[5:49] trusted autonomy solutions. This is a very complex
hard-coded behavior tree. Um there are more uh
[5:57] newer advancements like reinforcement learning and
state models, but at the end of the day, it's a
[6:02] very complex choose your own adventure book with
determined only a specific set of ways that it can
[6:08] end. Um next slide. So, why aren't we there yet,
right? Why aren't we as far along? Why aren't we,
[6:15] if we look at the art of the possible for
generative AI and non-physical autonomous systems?
[6:20] We're making leaps and bounds. Why are we not
moving as fast as we might think we should be for
[6:26] drones and autonomous systems? Uh, and there are
both technological and human constraints. Um, one
[6:33] something we call swap, right? Limited size and
limited power. If you think about the compute form
[6:40] factors that are required to run really advanced
AI systems, uh they're big, they're power hungry,
[6:47] they can be very loud and have a lot of, you know,
vibration or or need additional cooling systems.
[6:52] If we're providing that space on a 10-in
quadcopter that weighs no more than 10 lbs uh and
[6:57] providing that power, we are then reducing what
we can actually accomplish with their vehicle. So
[7:01] there is always a trade-off for drones really of
any form factor when we think about what compute
[7:06] do we want to put on board versus what mission
do we want to accomplish. Um and then there's
[7:11] the millisecond control loop. If we think about us
driving a car or a self-driving car when we're not
[7:17] really sure where we're going or the check engine
light comes on, we pull over, we stop the car, we
[7:21] read the manual, we get out Google, um we realign
ourselves on ways or Google Maps and that's fine.
[7:28] you can't really pull over a drone very easily.
Um, so that is really difficult. And then when
[7:34] you get to that third bullet, what happens when
the drone does absolutely have to pull over?
[7:38] It falls out of the sky. And the laws of physics,
especially gravity, are really immutable. And that
[7:43] can have really dire consequences generally in the
loss of aircraft, but also other really bad things
[7:47] can happen when very heavy aircraft or even a 10-
pound aircraft are falling out from 500 ft. Um,
[7:54] then we have difficulty in sensing. This one is
being overcome very rapidly, but it can be really
[7:58] hard to see clearly when you're 10,000 feet in
the air on a vibrating piece of carbon fiber,
[8:03] right? Not as easy as if you're whimo with a ton
of sensors on every angle of your vehicle. Um,
[8:10] and the next big thing is that it can't phone
home. So, we have beyond visual line of sight.
[8:15] We have denied integrated environments. We have
GPS denied environments. And these aren't true
[8:19] only for the military use case, right? If we're
doing critical infrastructure or we're inside a
[8:24] building or inside of a tunnel, we might not have
the ability to have that drone pick up the phone
[8:29] and call someone. Um, I think we all know
the self-driving car company that has a
[8:32] call center in the Philippines. So, the
car can actually call a person and say,
[8:37] "I need more context. I don't know what to
do." And that person isn't driving the car,
[8:41] but they are providing the context that car needs
to then continue to do what it's supposed to do.
[8:46] And even then, we still see issues, right? It's
parking on top of a fire hose in the middle of
[8:49] an active fire. it's accidentally
driving through a crime scene. Um,
[8:54] so this is a really difficult uh thing for drones
because we have that millisecond control loop
[9:00] and we can't always pick up the phone and
call. So context has to be really great in
[9:04] these situations. And then we get to the onetoone
problem. So this is a technology problem. It's an
[9:10] interface problem. It's a trust problem and it's a
legal problem. I think we're all familiar with the
[9:15] one drone to one operator, right? the person that
has the handheld remote control. Oh, can we go?
[9:22] Oh, that's okay. Um, I went to a leading
autonomy company's demonstration the other day
[9:27] and it was one autonomous air vehicle and three
operators. One to watch the C2 link, one to watch
[9:34] the autopilot in case the autonomy failed, and one
to watch the autonomy. So, we have three people
[9:40] for one autonomous drone. And we're kind of have
gone backwards. And I actually don't think this is
[9:45] a lack of ability of the autonomy solution.
I think it is a legal constraint that the FAA
[9:50] puts on these vehicles of what are your safety uh
what are your safety controls? And then also the
[9:56] company doesn't want to lose the air vehicle. They
want don't want to be responsible for a crash.
[10:00] They don't want to get their airspace taken away.
So now we're instead of one to one, we're 3:1.
[10:04] And there's a question of when are we ever going
to get to one to many in a way that we can trust.
[10:10] Next slide please. So, now that we're done with
the bad news, let's talk about the good news.
[10:13] These are some things that I have been really
excited about watching happen and evolve in
[10:18] this industry. Um, these aren't all brand new, but
they're newly fielded in ways that are trustworthy
[10:24] and are are actually being used today. So,
first we have the overtheair model updates,
[10:28] and these could be AI models, ATR models,
perception models, or even autonomy modules.
[10:34] So being able to click to deploy or
click to redeploy across a fielded fleet
[10:38] rather than having to land that aircraft and you
know plug your Ethernet cord into it and fiddle
[10:42] around with it. Uh this means that now autonomy
and AI on these vehicles can start evolving as
[10:49] fastly as the rest of the industry right as if
we were an enterprise or cloud-based solution.
[10:54] Um the next one translated commander intent or
operator's intent. So current autonomy solutions
[11:00] work a lot like that autopilot. I have to go into
the HMI, the interface. I have to click through
[11:06] what I want. I have to define the bounding
box, the no-fly zones, the no- look zones,
[11:10] uh ingress and egress. And the buttonology is
different for all the providers. And generally,
[11:16] the person who's saying I need this to happen
isn't also the person that knows how to use it.
[11:20] So, if we have a a very critical search and
rescue mission and I know I need to get out
[11:25] there rapidly, I still need to track down the
person that knows how to use the autonomy to do
[11:29] this and this and this and that. and that can take
a long time. So the really exciting advances in
[11:33] LLM, natural language understanding
and agentic workflows to get that
[11:38] please go search this area of the coastline for
the missing person to that autonomous vehicle
[11:42] at translating what I said into those autonomy
commands and having the vehicle go out and do.
[11:47] Um the scary part to me on this however is putting
those agentic agents onto the vehicle and then
[11:54] letting the vehicle determine outcome and intent.
Um so this we'll talk about this a little bit more
[11:58] in a minute. Um, but really great innovations, but
also I think it will take us a little bit of time
[12:02] to feel them in actuality. Um, and then the fun
cat and mouse game of countering the counter UAS.
[12:08] Every time we come up with a really great AI
solution to put on an air vehicle, uh, there's
[12:12] always a really great solution to counter the air
vehicle, whether it's lasers or nets or any other
[12:19] shotguns, who knows? Take them out of the sky. Um,
so it's been really fun to watch this evolve and I
[12:23] think this is one of the fastest moving segments
of this industry today. Um, and then to call out
[12:29] to a couple other things I think are really cool.
The advances in hardware, the neuromorphic trip
[12:33] chips are going to make it a lot easier for us to
put really great uh models onto the air vehicles.
[12:39] Uh, dual use directed energy has a lot of really
cool use cases. Uh, edgebased SLMs, as we continue
[12:44] to bring the model size down, they become much
more accessible for these types of vehicles. Um,
[12:51] and this is really important, right? Can we move
to the next slide, please? Um, what do we gain?
[12:54] Right? Uh, we aren't replacing the humans, right?
there. We all know humans are good at some things,
[13:00] machines are good at others, and let's kind
of like divide the lineup and let each do what
[13:03] they're good at. Um, little little segue, heaven
builds hydrogen-powered drones, and the hydrogen-
[13:09] powered drones give us a unique 8 to 10 hour
flight duration and very very quiet thermals. So,
[13:15] there's a lot of really unique mission sets that
you can do with a very quiet 10-hour air vehicle.
[13:21] Um, if we think about 10 hours, I for
one cannot pay attention for 10 hours.
[13:25] I think at 45 minutes of staring at a computer
screen looking for whatever I was looking for, I'm
[13:29] going to zone out. I'm going to get hungry. I'm
going to get distracted. I'm going to have to eat.
[13:33] I'm going have to go to the bathroom. Uh these
are things that machines don't have to do.
[13:37] Let's think about a bridge span inspection. If I'm
manually inspecting the bridge span of let's say
[13:42] the a three-mile bridge, I'm probably doing the
first bridge span really well. I'm very thorough.
[13:47] What happens when I get to the last bridge span?
Am I as thorough? Have I paid as much attention
[13:52] or have my eyes glazed over? Um, so by putting
autonomous and AI capabilities onto air vehicles
[14:00] that are very exquisite in their capability, we
are removing the cognitive burden from the human,
[14:05] which I think we're all familiar with. Um, and
most importantly, we're reducing exposure. We have
[14:09] very, very small areas that humans can't or don't
or won't go into. Um, think about a collapsing or
[14:15] burning building or rubble from a earthquake. We
also have active war zones or contested airspace
[14:21] where we don't want to expose people to those
environments. If we have good autonomy that we
[14:26] can trust, we no longer have to send people into
air. Um, so this is the real value proposition.
[14:32] Um, but we really have to get this right.
Can we go to the next slide, please? Um,
[14:37] there's what we risk and obviously there are a
lot of risks with AI. There are a lot of risks
[14:41] with autonomy. These are the ones that I think
are most important. One is the consequence.
[14:47] When an autonomous system makes a wrong decision
at machine speed in that very closed closed loop,
[14:52] generally we're not seeing it actively happen.
We're just seeing it after it happened.
[14:57] I don't know if everyone saw the recent video of
the humanoid robot collapsing on stage having to
[15:00] be drugged off by four or five people. Um, no one
saw that coming, but we certainly saw it when it
[15:05] happened. Um, if we think about an air vehicle
shuttling through the atmosphere, uh, we're
[15:11] probably only going to see it after it crashes. uh
very rarely do we see it happen as it's occurring.
[15:16] Then we have the attack surface, right? Every new
AI capability is a new AI attack surface. We have
[15:23] model poisoning, position spoofing, GPS spoofing.
We even have visual trickery of the ATR models
[15:29] that a lot of people are implementing. Um the more
autonomous the platform, the easier it could be to
[15:34] manipulate. So we really have to be careful with
that. Um and the one that I think about the most
[15:39] is the misplaced authority. So it isn't the risk
that we're letting the machines make the decision.
[15:46] It's that we let them make decisions we haven't
fully thought through. So we think about that
[15:49] autonomy gradient of uh the human with the remote
control and all the decision-making authorities
[15:55] with the human. And as we slide toward that
independent thinking, we are delegating all
[16:00] that authority to the machine. And we need
to be really careful and very intentional as
[16:04] we think about what is my intent, what is my
desired outcome, how do I want it to occur,
[16:10] and which one of those three things can I delegate
to the machine and trust that it will occur
[16:15] in the right and appropriate ways. Um, so it's
really about assigning the authority to act and
[16:21] where and that gradient [clears throat] you want
to give that authority. Um, next slide please.
[16:25] So why does this matter? Right? This is
a very fastmoving application space. Um,
[16:31] and the decisions we make right now, especially
about how when we delegate that authority and to
[16:36] what underlying autonomy and AI systems we're
delegating that authority to, the decisions
[16:41] we're making are going to impact the way this
industry evolves over the next 5 to 10 years.
[16:45] If we make really bad mistakes or move faster than
the trusted technology can move, I think it can
[16:51] set us back rather than help move us forward. Um,
so two really easy questions. Should drones be AI
[16:56] enabled? Of course. Should drones be autonomous?
of course, but I think the really hard problem is
[17:02] to which extent and where on that authority
or autonomy gradient we want to delegate. Um,
[17:08] so as someone specifically leveraging drones in
defense, military and homeland security missions,
[17:13] uh, the highest stakes missions in existence,
I don't believe AI has reached the point where
[17:18] we're ready to delegate that authority, that
critical decision- making and course of action
[17:22] to the drone. Um, let the human in the loop
determine the intent and the outcome and let the
[17:27] robot handle the execution. And that is all I have
for you guys today. I think we have maybe a couple
[17:33] minutes for questions. Okay, any questions from
the audience? So given that most of our technology
[17:39] problems aren't actually technology problems and
given the humans involved in the geopolitical
[17:44] situations that we're in right now, what do
you give the odds that we do do this right? I'm
[17:49] sorry. Given the odds that we do what? That we do
this cor that we do this well and not recklessly.
[17:56] I think that's really interesting. In the panel
earlier, it was brought up the difference between
[18:00] what we do at war and what we do at peace.
Um, and I think we're seeing a lot of really
[18:05] innovative things happening in countries where
there is war. Israel and Iran, Ukraine and Russia,
[18:10] we're seeing a lot of really, really great
innovations. When those come to the United States,
[18:15] what I've specifically seen is we are as a defense
industrial base reticent to adopt them because
[18:22] do we trust it? Do we not trust it? Do we need to
do that right now? Do we have more time to think
[18:27] through the implications of that? Um, so that's
a great question. I think we are intentionally
[18:34] taking our time right now, but there is always
the chance that those risks and the need to
[18:39] take those risks overcome our ability to take
our time and develop these trusted algorithms.
[18:47] That is our time. Let's clap it up for Sammy
Hamilton. Thank you so much. Thank you guys.
