Worlds Uses Physical AI to Cut Enterprise Check-In Times to 10 Seconds
Ai4
Physical AI delivers measurable enterprise returns where generative language models stall by converting live sensor feeds into spatial operational baselines that automate industrial workflows in real time.
## Background
Published on YouTube by the Ai4 channel on August 26, 2025, this presentation features Dave Copps, CEO and Co-founder of Worlds, examining how Physical AI transforms enterprise operations. According to the description, the session explores how organizations can move beyond AI buzz to deliver real-time operational insights, increased productivity, enhanced safety, and measurable ROI at scale.
## TL;DR
While back-office generative AI often struggles to demonstrate tangible financial returns, Physical AI embeds artificial intelligence directly into industrial environments to deliver concrete, A/B-testable operational metrics. The Worlds platform ingests existing camera and sensor feeds into Large World Models to automatically map an operational baseline known as the 'Pattern of Life,' completely replacing tedious manual data labeling with instant automated clustering. Enterprise deployments demonstrate dramatic results, including reducing trucking terminal gate check-ins from seven minutes to ten seconds, detecting steel mill defects to save thousands per incident, and catching million-dollar pipeline leaks. By combining Forward Deployed Engineering teams with multi-sensor contextual fusion, organizations achieve two-to-four-week rollouts while driving enterprise false alarms down to near zero.
## Atmosphere & style
The presentation is brisk, pragmatic, and intensely solutions-oriented, deliberately cutting through enterprise disillusionment with generative AI hype to focus on tangible physical engineering. Dave and Chris establish a dynamic tag-team rhythm, pivoting effortlessly between high-level market context and granular, live platform demonstrations spanning tarmac tracking to heavy steel cauldrons. The experience feels grounded and authoritative, offering viewers a direct look at the industrial frontier where software interfaces give way to real-world sensors, physical actions, and concrete operational returns.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=4s) Unlike back-office generative AI where ROI is often subjective and difficult to quantify, Physical AI delivers clear, measurable returns by running direct A/B testing on physical throughput, safety, and operational efficiency.
- [01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) The user interface for Physical AI moves beyond computer monitors into the physical environment itself through vibrating alerts, automated signaling, and dynamically adapting facilities.
- [03:01](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=181s) While Large Language Models face data saturation limits with digital text, Physical AI unlocks an exponentially larger, untapped training dataset continuously generated by the physical world.
- [04:45](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=285s) Worlds structures raw enterprise video and sensor streams into Large World Models based on the military concept of 'Pattern of Life'—establishing a spatial baseline of who, what, where, and when.
- [05:21](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=321s) The platform replaces slow, manual bounding-box data labeling with an automated bottom-up visual indexing and clustering engine, cutting model training time from weeks to minutes or hours.
- [08:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=493s) Automated physical AI surveillance delivers a 10x capacity expansion in monitoring coverage while reducing operational false alarms by 80% to 90%.
- [11:34](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=694s) Enterprise deployments demonstrate quantifiable operational impact, including reducing trucking terminal gate check-ins from 7 minutes to 10 seconds and averting million-dollar pipeline containment liabilities.
- [14:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=853s) Deploying Forward Deployed Engineering teams compresses enterprise implementation cycles from six months down to two to four weeks.
- [15:34](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=934s) Pairing single-point hardware sensors with visual camera confirmation provides contextual situational intelligence, virtually eliminating false alarms that previously numbered up to 100,000 per month.
- [16:51](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1011s) Physical AI overcomes the blind spots of legacy computer vision by adopting anomaly and aberration detection to catch unexpected hazards without requiring pre-labeled training for every failure mode.
- [18:16](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1096s) Unified multi-sensor fusion synchronizes camera feeds, GPS telemetry, and IoT signals into a cohesive spatial-temporal model capable of self-supervised operational routine recognition.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=4s) Beyond Failed AI Projects: The Tangible ROI of Physical AI
Following a preceding speaker who emphasized that the vast majority of enterprise AI initiatives fail, Dave Copps introduces Worlds' approach as an antidote to this trend. While the past couple of years have seen enterprise hype dominated by generative AI, proving return on investment in the back office has remained frustratingly elusive. Generative AI tools often yield subjective, abstract productivity gains where employees feel more productive, but organizations struggle to quantify the exact financial impact.
Physical AI shifts the paradigm by taking artificial intelligence out of back-office software and embedding it directly into the physical environments of enterprise operations. In physical environments, proving value is concrete: organizations can conduct straightforward A/B testing by comparing baseline operational metrics before deployment directly against post-deployment performance in areas like throughput, efficiency, and workplace safety.
[01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) The Physical AI Awakening: NVIDIA GTC and Beyond the Back Office
While physical AI broadly encompasses autonomous vehicles and humanoid robotics, the immediate commercial opportunity centers on creating intelligent environments that augment enterprise workforces. Instead of confining AI's user interface to a "pane of glass" on a computer monitor, the user interface of physical AI is the physical world itself—manifesting through vibrating mobile alerts, automated physical signaling, and operational spaces that dynamically adapt to real-time events.
```mermaid
flowchart TD
accTitle: Physical AI Transformation Architecture
subgraph SENSING["1. Physical Ingestion"]
A["Existing CCTV & Video Cameras"]
B["Thermal & IoT Sensors"]
end
subgraph WORLD_MODEL["2. Worlds Platform (Large World Model)"]
C["Bottom-Up Auto-Indexing & Clustering"]
D["Pattern of Life Engine\n('Who/What, Where, When')"]
E["Live Spatial Data Stream"]
end
A --> C
B --> C
C --> D
D --> E
E --> F
F --> G
```
The inflection point for this market arrived in March at NVIDIA GTC in San Jose. In front of an arena of 20,000 attendees, Jensen Huang closed his two-hour keynote with an industry trajectory slide placing physical AI at the upper right, declaring that it is time for AI to understand the laws of physics and predict physical occurrences. This perspective was quickly reinforced by Julie Sweet, CEO of Accenture, who declared it "the year of physical AI."
- Physical AI takes AI out of back-office software and embeds it into physical enterprise operations.
- The interface moves from digital screens to real-world actions: buzzing worker devices, warning lights, and automated facility responses.
- NVIDIA GTC in San Jose marked the mainstream turning point for physical AI's enterprise adoption.
[03:01](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=181s) Unlocking World Data: Why Language Models Aren't Enough
The timing of physical AI aligns directly with the structural limits currently facing Large Language Models. AI developers have essentially ingested the entirety of available digital document training data. While human culture continues to generate text, the textual training corpus is largely saturated. In contrast, an exponentially larger volume of untapped training data remains locked in the physical world.
The most effective way for an AI system to understand reality is not merely reading about it in documents, but experiencing it continuously. By connecting directly to an organization's existing sensor grids and security cameras, AI can observe, interpret, and learn from physical phenomena firsthand. Momentum has surged rapidly over recent months, with physical and agentic AI being recognized across mainstream financial and tech coverage as multi-trillion-dollar market opportunities.
[04:45](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=285s) Large World Models and the Military Baseline of 'Pattern of Life'
Rather than training Large Language Models on text, Worlds builds "Large World Models" by ingesting real-time feeds from existing enterprise cameras and IoT sensors, transforming raw visual footage into a structured, live spatial data stream. This data stream captures what is being sensed, precisely where it is located, and when it occurs.
To make sense of physical spaces, Worlds adopts a foundational operational concept from the military known as "Pattern of Life." Before deploying complex workflows, the platform connects to existing sensor arrays to establish a baseline understanding of normal operational movements: who is where and when, and what assets are where and when. Once this baseline data stream is established, organizations can deploy autonomous agents and specialized software applications on top of the platform to address specific industrial problems.
- Live spatial data serves as the new foundational raw material for digital transformation in physical facilities.
- "Pattern of Life" provides the non-negotiable operational baseline by tracking spatial occupancy and asset flow over time.
- Sensor data is converted into actionable, structured data streams without requiring dedicated hardware overhauls.
[05:21](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=321s) Architecture of a Physical AI Platform: Flipping Data Labeling on Its Head
Traditional computer vision training relies on a labor-intensive, top-down manual process where human annotators spend hundreds of hours manually drawing bounding boxes around objects—a model popularized by companies like Scale AI. Worlds discarded this approach, replacing it with a fully automated, bottom-up indexing system that trains computer vision models in minutes to hours rather than weeks.
When cameras are connected to the platform, the system automatically indexes every visible object without human intervention. Operators are provided with an intuitive visual search and clustering interface where selecting an object and clicking "more like this" rapidly aggregates matching visual entities across lighting conditions, shapes, and motion profiles. Once trained in a single day, the platform runs real-time inference on live sensor feeds, enabling Forward Deployed Engineering (FDE) teams to deploy custom agentic solutions on-site within two to three weeks.
[08:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=493s) From Airport Tarmacs to Hospital Security: Instant Training and Anomaly Detection
Chris takes over to demonstrate the practical application of the platform, beginning with an aircraft-indexing demo inspired by former FedEx CIO Ken Spangler using footage from Amsterdam Airport Schiphol. After raw video is ingested, selecting a single aircraft and utilizing the clustering feature acts like a visual magnet, automatically aggregating 1,300 matching aircraft examples across varying angles, shapes, and movement dynamics within a couple of minutes.
Turning to enterprise deployments, Chris illustrates how physical AI transforms facility security in environments like hospitals. A human security team cannot effectively monitor hundreds of live camera feeds simultaneously, especially across long shifts or high-temperature outdoor conditions where fatigue inevitably sets in. By training the AI on a facility's floor plan to distinguish between doctors, nurses, patients, and visitors, the system monitors all cameras concurrently, detects perimeter breaches and abnormal movements, and presents analysts with a clean, color-coded spatial interface.
- A visual clustering tool can assemble 1,300 valid training examples from raw video feeds within minutes.
- Hospital security deployments automate perimeter breach detection and staff/visitor categorization.
- Facilities achieve a 10x capacity improvement in monitoring coverage while driving an 80% to 90% reduction in false alarms.
subgraph DECISION["2. Physical AI Engine"]
D{"3-Way Match\nVerified?"}
end
subgraph ACTION["3. Terminal Flow"]
E["Instant Clearance\n(7 Min → 10 Sec)"]
F["Hold for Inspection"]
end
A --> D
B --> D
C --> D
D -- Yes --> E
D -- No --> F
```
In logistics, Worlds is automating 18 national terminals for a top-five US trucking company. At the guard gate, the system reads tractor IDs, scans container serial numbers, and authenticates carrier logos simultaneously. If all three match, the truck passes through automatically. This reduces gate check-in duration from seven minutes down to just ten seconds—a 92% productivity boost that prevents roadside queueing, ensures drivers (who are paid by the mile, not the hour) are not delayed, and guarantees tractors do not haul incorrect payloads.
In heavy manufacturing, a top-five steel producer uses the platform across three distinct operational areas: optimizing cut precision to maximize yield, verifying that rolled steel is straight to catch structural defects early (saving $2,300 per detected incident), and utilizing thermal imaging to monitor molten steel cauldrons to prevent catastrophic overfilling while protecting nearby personnel. Finally, for a major pipeline operator, Worlds combines thermal imaging with change-detection algorithms to identify minute pipeline leaks in real time, preventing cumulative environmental damage and remediation costs that easily exceed $1 million.
- **Trucking Logistics:** Terminal check-in times cut from 7 minutes to 10 seconds (92% productivity gain) across 18 enterprise terminals.
- **Steel Manufacturing:** Early structural defect identification saves $2,300 per incident alongside thermal cauldron safety monitoring.
- **Pipeline Infrastructure:** Combined thermal and change-detection algorithms spot micro-leaks early, averting million-dollar containment and environmental liabilities.
[14:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=853s) Forward Deployed Engineering and Rapid Time-to-Value
Addressing how quickly enterprise deployments deliver returns, Dave explains that their answer has completely transformed over the past year. In January, Worlds established a Forward Deployed Engineering (FDE) team designed to embed directly with customers—either on-site or virtually—to compress implementation cycles and achieve rapid ROI. Workflows that previously took six months are now routinely completed in two to four weeks.
The FDEs begin by establishing a baseline "pattern of life"—converting the physical dynamics of people, assets, and operations into structured spatial-temporal data—and then asking the customer directly: *"What problem are you trying to solve?"* Leveraging their core platform, APIs, and SDKs, the team builds tailored applications on top of that baseline.
As a concrete example, Worlds partnered with one of the world's largest engine manufacturers. In just four weeks, they moved from zero to a fully operational, live health and safety dashboard spanning the entire plant floor, identifying hazardous occurrences in real time.
```mermaid
flowchart LR
accTitle: Forward Deployed Engineering Workflow
subgraph DataLayer["1. Establish Baseline"]
A["Physical Site Activity"] --> B["Pattern of Life Data<br/>(Who / What / Where / When)"]
end
subgraph SolutionLayer["2. Solution Build"]
B --> C["FDE Team & Customer Problem Definition"]
C --> D["Worlds Core AI, APIs & SDK"]
end
subgraph OutcomeLayer["3. Deployment"]
D --> E["Live Real-Time Dashboard<br/>(2–4 Week Time-to-Value)"]
end
```
- **Deployment Velocity**: Enterprise implementation timelines dropped from six months down to two to four weeks through the introduction of the forward-deployed engineering model.
- **Engine Plant Case Study**: Delivered an end-to-end plant-wide safety monitoring dashboard showing live incidents for a major engine manufacturer within a month.
---
[15:34](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=934s) Model Accuracy and Eliminating False Positives Through Sensor Context
Addressing model performance and verification, the speakers point out that their AI quality is rigorously quantified across high-stakes commercial environments. In partner deployments across casinos, for instance, Worlds models analyze live gameplay by tracking and validating every single gaming chip and table action in real time, measuring precise model efficacy against real-world performance.
A major flaw in traditional physical AI systems has been their susceptibility to false alarms—often because enterprises rely on isolated single-point sensors. Organizations frequently report dealing with upwards of 100,000 false positives per month. Worlds resolves this issue by constructing contextual intelligence through multi-sensor pairing.
```mermaid
flowchart TD
accTitle: Sensor Fusion Context Flowchart
A["Door Sensor Triggered (Open)"] --> B{"Camera Visual Verification"}
B -->|"Truck Present at Bay"| C["Status: Open & Occupied<br/>(Suppress False Alarm)"]
B -->|"No Truck / Truck Departs"| D["Status: Open & Unoccupied<br/>(Trigger Security Alarm)"]
```
If a magnetic door sensor indicates an open bay while an overhead camera confirms a loading truck is docked, the system classifies the bay as "open and occupied" and suppresses the alarm. The second the truck pulls away and the door remains open, the state changes to "open and unoccupied," immediately generating a genuine security alert. By combining sensor feeds into coherent situational context, false positives drop essentially to zero.
- **Casino Gameplay Tracking**: High-precision computer vision models track micro-actions and gaming chips at casino tables to quantify AI quality.
- **False Alarm Elimination**: Contextual sensor validation resolves operational noise, driving down enterprise false-positive volumes that historically reached 100,000 per month.
---
[16:51](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1011s) Beyond Legacy Computer Vision: Drone Inspections and Agriculture
The discussion broadens into aerial inspections and complex open-world environments like agriculture. Worlds is utilizing drone flyovers for infrastructure tasks, such as automated visual inspection and condition monitoring along energy pipelines.
Dave highlights a fundamental limitation of legacy computer vision: *"It only sees what you train it to see."* In traditional systems, if a model trained specifically to detect children in a crosswalk encounters an elephant walking across the street, the AI completely ignores the elephant. Physical AI overcomes this through aberration and anomaly detection.
In agricultural environments, instead of attempting to catalog every potential pest or defect, models can be trained on the baseline appearance of healthy fruit or foliage and instructed to flag any anomalous deviation or browning. Furthermore, true spatial-temporal sensor fusion allows systems to ingest and correlate disparate data streams—such as live field cameras, drone imagery, weather feeds, and soil moisture sensors—synchronized across space and time to produce holistic crop intelligence.
- **Infrastructure & Drones**: Drones paired with Physical AI automate pipeline inspection by evaluating conditions from aerial passes.
- **Anomaly Detection Paradigm**: Moving away from rigid classification toward aberration detection allows models to flag unknown issues without requiring pre-labeled training for every failure mode.
---
[18:16](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1096s) Full Multi-Sensor Fusion and Pattern of Life Modeling
While cameras remain the team's favorite sensor due to their high information density—enabling the system to determine not just object identity but exact operational state (e.g., whether a container or space is full or empty)—Worlds operates as a full-spectrum sensor fusion platform.
The architecture ingests diverse telemetry, including vehicle GPS trackers, personnel badges, and IoT hardware, overlaying a single unified AI layer across all inputs to model real-world physics, trajectories, and interactions:
- **Unified Physical Modeling**: Synchronizes spatial coordinates, physical trajectories, and time series data across cameras, GPS, and personal trackers.
- **Self-Supervised Pattern Recognition**: Automatically aggregates repetitive spatial routines—such as an aircraft landing, taxiing, and docking at an airport gate—and classifies them into cohesive operational milestones (e.g., an "arrival event") without requiring manual rule-writing for every sub-step.
The session concludes with appreciation for the audience as the presenters wrap up their walkthrough of Physical AI in commercial operations.
## Notable quotes
> [01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) it is time for AI to understand the laws of physics and predict physical occurrences — Jensen Huang
> [01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) the year of physical AI — Julie Sweet
> [04:45](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=285s) Pattern of Life — Dave Copps
> [05:21](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=321s) more like this
> [14:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=853s) What problem are you trying to solve?
> [15:34](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=934s) open and occupied
> [15:34](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=934s) open and unoccupied
> [16:51](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1011s) It only sees what you train it to see. — Dave Copps
## Video description
Dave Copps, CEO & Co-founder, Worlds
Abstracts
Join us to explore how Physical AI is transforming the daily operations of enterprises worldwide, helping them move beyond AI buzz to delivering real-time insights and results that drive increased productivity and enhance safety, while generating measurable ROI at scale.
Dave Copps and Chris of Worlds argue that while generative AI often produces intangible back-office gains, Physical AI delivers verifiable return on investment by embedding spatial intelligence directly into enterprise operations.
Moving Beyond Back-Office Language Models (0:04)
Enterprise adoption of generative AI has frequently stalled due to difficulties in proving financial return, yielding subjective productivity impressions rather than concrete operational gains. Physical AI shifts models from computer screens into industrial environments, allowing companies to evaluate performance through direct A/B testing of metrics such as throughput, safety, and operational efficiency.
This shift coincides with structural limits in large language models, where the digital document corpus has become largely saturated. Dave Copps highlights that an exponentially larger volume of operational data remains unindexed in the physical world. Industry leaders, including NVIDIA CEO Jensen Huang and Accenture CEO Julie Sweet, have pointed to this transition as a critical inflection point where AI systems must learn physical laws and predict real-world events.
Automated Spatial Indexing and Baseline Modeling (4:45)
Traditional computer vision relies heavily on labor-intensive manual labeling, requiring human annotators to draw bounding boxes over hundreds of hours. Worlds replaces this with an automated bottom-up indexing system: connecting existing CCTV cameras and IoT sensors allows the platform to index visible objects without upfront manual annotation, enabling visual clustering tools to aggregate over 1,300 matching entity examples within minutes.
To interpret facilities systematically, the platform adopts the military concept of 'Pattern of Life.' By tracking spatial occupancy and movement over time, the system establishes an operational baseline of who and what is present across specific spaces and times. This live spatial data stream serves as the foundational layer upon which specialized software agents and workflows are deployed, avoiding the need for dedicated hardware overhauls.
Measurable Field Outcomes in Enterprise Deployments (11:34)
Enterprise implementations demonstrate quantifiable returns across logistics and heavy industry. In logistics, Worlds automated guard gates across 18 national terminals for a top-five US trucking company. By simultaneously reading tractor IDs, container serial numbers, and carrier logos for a three-way match, gate check-in duration dropped from seven minutes to ten seconds—a 92% productivity improvement that prevents roadside congestion and dispatch delays.
In manufacturing, a top-five steel producer uses the platform across three operational areas: optimizing cut precision, detecting rolled-steel structural defects early to save $2,300 per incident, and monitoring molten steel cauldrons with thermal imaging to avert overfilling. Similarly, a major pipeline operator pairs thermal imaging with change-detection algorithms to detect micro-leaks in real time, preventing environmental damage and remediation costs that exceed $1 million.
Sensor Fusion, Anomaly Detection, and Rapid Deployment (14:13)
Legacy physical security and monitoring architectures often generate severe operational noise, with enterprises reporting upwards of 100,000 false positives monthly. Worlds counters this by fusing multiple sensor feeds into situational context; for example, cross-referencing a magnetic door trigger with overhead camera footage confirms whether an open loading bay is actively occupied by a truck, suppressing false alarms unless an unauthorized vacancy occurs.
Furthermore, the system transitions from rigid object classification toward anomaly and aberration detection, enabling models to flag deviations—such as infrastructure damage or agricultural crop browning—without requiring pre-labeled training for every possible defect. Deployment timelines have also compressed: by embedding Forward Deployed Engineering teams with clients, standard enterprise implementations dropped from six months to between two and four weeks, exemplified by a plant-wide safety dashboard deployed for a major engine manufacturer within a month.
The presenters conclude that uniting existing sensor grids with real-time spatial modeling allows enterprises to transition computer vision from passive surveillance into automated operational workflows. The commercial viability of this approach hinges on sustaining low false-positive rates and executing rapid, customized on-site deployments without demanding manual retraining or costly infrastructure replacement.
## Background
该视频由 YouTube 频道 Ai4 于 2025 年 8 月 26 日发布，由 Worlds 联合创始人兼 CEO Dave Copps 与团队成员 Chris 共同演讲。内容聚焦于物理 AI（Physical AI）如何帮助全球企业走出生成式 AI 的投资回报率迷雾，利用现有传感器网络构建智能环境，在提升安全与生产力的同时实现切实可衡量的规模化 ROI。
## TL;DR
演讲指出生成式 AI 在企业端常面临 ROI 模糊难以量化的痛点，而物理 AI 通过将“真实世界本身作为 UI”，将海量物理运营数据转化为动态空间数据流。Worlds 依托“大世界模型”与军方级“生活模式”（Pattern of Life）基线，颠覆了传统耗时的人工标注流程，并通过前线部署工程团队（FDE）在 2 至 4 周内完成方案上线。分享中展示了医院安防、货运物流闸口、钢铁制造与油气管网泄漏等多个落地案例，证明通过时空同步的多传感器融合与异常检测算法，不仅能大幅降低误报率，更能带来直接且可验证的硬性经济效益。
## Atmosphere & style
整场演讲节奏紧凑、干货密集且极具说服力，展现了从前沿技术构想到重工业落地无缝衔接的务实风格。两位讲者通过现场高频的软件交互演示与硬核业务数据（如单次纠错节省 2,300 美元、闸口放行提速 92% 等），直接扫清了传统 AI 概念演讲的悬浮感。现场氛围专业而笃定，将观众从对虚拟文本模型的探讨迅速拉回轰鸣的工厂车间与物流码头，带来强烈的工业智能化临场感。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=4s) 生成式 AI 的投资回报率往往主观且难以捉摸，而物理 AI 直接切入真实运营现场，通过直观的前后 A/B 测试提供清晰、可量化的刚性 ROI。
- [01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) 物理 AI 将现实世界本身作为交互界面（UI），核心在于将 AI 能力直接嵌入企业物理运营环境，赋予其自动化并实时感知测量物理流程的能力。
- [03:15](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=195s) 大语言模型面临人类数字化文本数据耗尽的瓶颈，而物理世界蕴藏着更庞大的潜在数据，AI 唯有通过摄像头与传感器亲身体验与学习物理规律。
- [04:45](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=285s) Worlds 采用“大世界模型”架构，以“生活模式”（Pattern of Life）为核心基线，在数据层面理清人与物在时空中的常态分布规律。
- [05:50](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=350s) 物理 AI 平台颠覆了传统耗时费力的人工标注流程，利用自下而上的无监督特征索引与小样本聚类交互，将模型训练周期从数周压缩至数小时。
- [08:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=493s) 在大型综合医院安防场景中，物理 AI 依靠实体分级识别与生活模式异常预警，将人员处置能力提升近 10 倍，并将误报率骤降 80% 至 90%。
- [11:34](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=694s) 在大型物流枢纽场景中，物理 AI 自动比对牵引车、集装箱号与货运标识，使进场核验提速 92%（由 7 分钟缩至 10 秒），杜绝货柜错拉事故。
- [12:30](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=750s) 在钢铁制造与油气管网场景中，高精度尺寸监控、高温熔炉液位检测与热成像微小泄漏检测，可直接为企业挽回数千至百万美元的损耗与罚款。
- [14:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=853s) 通过设立前线部署工程团队（FDE）并结合成熟 API/SDK 架构，Worlds 将传统长达 6 个月的 AI 落地周期大幅缩短至 2 到 4 周。
- [16:02](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=962s) 消除工业现场海量警报疲劳的关键在于“通过传感器组合构建上下文”，将视觉画面与门禁等物理传感器交叉验证，大幅根除单一传感器误报。
- [16:48](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1008s) 突破传统视觉“只能识别预设训练目标”的缺陷，现代物理 AI 转向异常检测（Aberration Detection）机制，通过学习健康常态即可敏锐捕捉各类异常。
- [18:16](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1096s) 物理 AI 的终极形态是时空同步的全传感器融合，系统能够自主内化物理时空运动规律并聚合识别连续事件，无需繁琐硬编码规则。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=4s) 破除生成式 AI 的投资回报率迷雾：物理 AI 的崛起与落地
Dave Copps 登台时顺着上一位演讲者关于“大多数 AI 项目正在走向失败”的论调切入，直言这恰恰为本次分享提供了绝佳的铺垫。过去几年间，企业界在评估生成式 AI（Gen AI）的投资回报率（ROI）时普遍陷入了泥潭。生成式 AI 带来的产出往往是“模糊且难以捉摸的”（fuzzy and intangible），人们主观上觉得生产力似乎提高了、事情似乎变好了，但在硬性指标上却根本说不清背后的机理、价值几何以及具体原因。
物理 AI（Physical AI）的出现打破了这种抽象与主观。它的核心是把人工智能从传统的“后台办公室”（back office）解放出来，直接带入现实运营环境。当 AI 接触物理世界时，ROI 变得前所未有地直观可见——企业能够清楚地观察到系统运行的效果，甚至可以像软件一样进行严谨的 A/B 测试：“这是引入物理 AI 之前的作业环境，那是引入物理 AI 之后的作业环境”。两者对照之下，生产率、运营效率以及工作场所安全性的提升都一目了然。
- **生成式 AI 的痛点**：成效高度主观、抽象，缺乏可量化、可验证的刚性 ROI 验证机制。
- **物理 AI 的可衡量性**：将现实环境转化为可测量的数据，通过前后对照（A/B Testing）清晰呈现效率与安全维度的质变。
---
### [01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) 什么是物理 AI：以真实世界为 UI 的智能环境
在广义语境下，物理 AI 常让人联想到自动驾驶汽车或人形机器人，但 Worlds 聚焦的重点在于“智能环境”（Intelligent Environments）。物理 AI 的实质，是将 AI 能力直接嵌入全球大型企业的物理运营现场，赋予它们实时测量、分析并构建自动化物理流程的能力。在传统的后台软件中，用户界面（UI）无非是一块冰冷的玻璃屏幕（"the UI is a pane of glass"）；而在物理 AI 的世界里，"UI 就是现实世界本身"——它体现在员工口袋里震动的手机、现场闪烁的警示灯，以及整套环境如行云流水般自动协同运转的细节之中。
这一波技术浪潮的“发令枪”在同年 3 月正式打响。在加利福尼亚州圣何塞举办的英伟达 GTC 大会上，黄仁勋（Jensen Huang）面向体育馆内两万名参会者进行了长达两小时的开幕演讲。在展示技术演进路径的关键 PPT 上，右上角赫然标注着“物理 AI”。黄仁勋断言，人类终于迎来了与“理解物理定律”的 AI 进行交互的时代，这种 AI 不仅能理解当下物理空间发生的一切，更能对未来态势做出精准预测。埃森哲（Accenture）全球 CEO 随后也明确表态，宣布今年是“物理 AI 之年”。
这场范式转移的底层逻辑在于大语言模型（LLM）的训练数据瓶颈。语言模型几乎已经吃透了全人类现存的所有数字化文档，在很大程度上“耗尽了训练数据”。然而，封存在物理世界中的潜在数据量，远比人类历史上所有文档所能记载的信息要庞大得多。想要让 AI 真正理解现实世界，仅靠阅读文本是远远不够的，最好的方式是像人类一样去“亲身体验”。通过接入企业现有的摄像头与传感器网络，AI 得以第一次以人类的感知维度来观察、体验并学习现实世界。
---
### [04:45](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=285s) 大世界模型与“生活模式”（Pattern of Life）基线
Worlds 摒弃了大语言模型路线，转向构建“大世界模型”（Large World Models）。该架构直接连通企业现有的传感器网络，将捕获到的海量现实画面与信号实时转化为动态空间数据流（Live Spatial Data）。一旦驱动企业价值的核心物理活动全部完成数字化，上层的具体应用与 Agent 便能在此基础上发挥出巨大威力。
```mermaid
flowchart TD
accTitle: Worlds 物理 AI 端到端架构流程
subgraph S1["感知与训练层"]
A["现有摄像头与传感器"] --> B["自下而上自动特征索引"]
B --> C["小样本聚类与秒级标注"]
end
subgraph S2["核心基线建模"]
C --> D["生活模式 (Pattern of Life)<br/>谁在何时何地/物在何时何地"]
D --> E["动态空间数据流 (Live Spatial Data)"]
end
subgraph S3["执行与交付层"]
E --> F["空间智能引擎 / Agent 架构"]
F --> G["前向工程团队 (FTE)<br/>2-3周快速落地可衡量 ROI"]
end
```
在市场落地过程中，Worlds 借鉴了军方的核心概念——“生活模式”（Pattern of Life）。无论进入哪个行业，团队首要确立的基线都是理清“谁在何时出现在何地，物在何时处于何处”。唯有把这种动态规律作为底层数据牢牢锁定，才能在此基线之上叠加高级算法，解决具体的业务难题。
为了支撑这套体系，Worlds 打造了一套端到端的物理 AI 平台，并在模型训练机制上做出了彻底颠覆：
- **颠覆传统人工标注**：传统计算机视觉极度依赖 Scale.com 这类数据标注公司，动辄耗费全球数十万人耗时数周在屏幕上人工画框；Worlds 转向全自动、自下而上的架构，摄像头通电后即可自动对环境要素进行无监督索引与特征提取。
- **极速交互与标注**：用户通过类似搜索引擎的聚类界面勾选关键目标并点击“查找类似项”（More like this），即可瞬间完成特定场景的标注，将原本数周的训练周期压缩至数小时乃至当天完成。
- **快速商业化变现**：依托确立的生活模式底座，Worlds 部署前向工程团队（Forward Engineering Team, FTE）进驻客户现场，通常在两到三周内即可跑通闭环并交付明确的 ROI。
---
### [08:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=493s) 现场案例实测：从极速模型训练到医院安全体系
Chris 接过话筒，以极快的节奏拆解了多个真实的落地案例。他首先展示了受前联邦快递（FedEx）CIO Ken Spangler 启发的模型训练演示。虽然出于保密无法直接展示联邦快递的现场画面，但团队使用阿姆斯特丹史基浦机场（Schiphol Airport）的公开视频流进行了还原。系统载入原始视频后自动框选出所有移动实体，操作人员选中一架飞机并点击“查找类似项”，算法便像磁铁一样将所有具有相似形状、颜色、尺寸、纹理乃至运动轨迹的目标一并吸取。在台上短短几分钟的实时交互中，系统就精准捕获了 1,300 个训练样本，彻底颠覆了过往耗费数百小时手工画框的传统方式。
随后，Chris 演示了物理 AI 在某大型综合医院安全体系中的部署。在复杂的安防环境下，单靠人类肉眼盯梢数百个摄像头显示屏是不切实际的，即便安保监控室坐满员工也无济于事，更不用说安保人员在 110 华氏度（约 43.3 摄氏度）的酷暑户外巡逻极易产生生理疲惫与疏漏——“AI 永远不会感到疲倦，它不知道什么是疲劳”。
在医院系统的数字界面上，物理 AI 直接投影在建筑平面图上运行：
- **智能实体分级识别**：AI 负责消化原始视频流，并转化为分析师视角的简洁图层。系统用不同颜色编码实时区分医生、护士、患者、访客（蓝色标点）以及正在巡逻的专业安保人员（红色标点）。
- **生活模式与异常预警**：系统通过建立院区行为常态基线，仅在发生翻越围栏、越界闯入或反常滞留等异常事件时触发告警。
- **业务收益**：现场安保人员的日常处置能力直接提升了近 10 倍（10x improvement），而传统安防系统中最让运营团队头痛的误报率，在部署物理 AI 后骤降了 80% 至 90%，带来了巨大的隐性人力成本缩减。
---
### [11:34](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=694s) 工业与物流场景落地：卡车货运、钢铁制造与管道泄漏检测
Chris 紧接着展示了三大重工业与物流场景的具体经济效益，证明物理 AI 正在传统实体产业中兑现真金白银：
在一家全美排名前五的大型货运物流企业中，Worlds 正在为其遍布全美的 18 个陆运枢纽站部署物理 AI。系统目前优先架设在园区警卫岗亭关口，后续将延伸至调度货场与维保设施。过去货车进出需要繁琐的人工登记与核验，常导致货车在场外公路大排长龙；如今，门禁处的物理 AI 能够在车辆行进间自动读取牵引车特征、识别集装箱箱号代码，并交叉比对车身货运标识。三项数据比对无误后，闸口直接抬杆放行。
- **通关提速 92%**：单车进场核验时间从原来的 7 分钟大幅压缩至仅需 10 秒。
- **驾驶员满意度与防错**：货车司机是按里程而非按工时计酬，长期拥堵等待是离职与不满的主要来源；同时自动化校验彻底杜绝了司机“拉错货柜半路才发觉”的致命调度失误。
在全美排名前五的一家大型钢铁制造厂内，物理 AI 同时跑通了三大业务用例。通常客户会有一个最具吸引力的“杀手级”用例，并连带部署多个伴生用例：
- **高精度切割与缺陷拦截**：系统实时监测钢材切割精度与平直度，及早截断缺陷工序。每一次提前捕获钢材变形或缺陷，都能直接为工厂挽回 2,300 美元的物料浪费与返工损耗——单凭这一项省下的成本，就足以覆盖整套系统的采购与部署开销。
- **高温熔炉溢流监控**：利用红外热成像摄像头实时“穿透”大型铁水钢包包体表面，监测液位高度，严防铁水超载溢出。
- **危险区域人员防护**：通过视觉划定禁区，实时监控并阻止工人踏入潜在的高温机械伤害区域。
最后展示的是一家全美大型油气管道运输商的应用场景。针对隐患极大的长输管网微小泄漏，Worlds 将高灵敏度热成像视觉与变化检测算法（Change Detection）相融合。热成像画面中的微弱温差与形态异常一旦被算法勾稽确认，便会立刻定位并发出告警。这种极其微小的泄漏若长期潜伏未被发现，造成的物料损失与环境整治罚款往往在数月内就会超过 100 万美元，而物理 AI 的早期干预直接将重大事故扼杀在了萌芽阶段。
### [14:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=853s) 前线部署工程团队（FDE）与快速 ROI 交付
如果在一年前被问及技术落地周期，Worlds 团队给出的答案会截然不同。为了打破传统 AI 项目周期漫长的痛点，公司在今年 1 月组建了“前线部署工程团队”（Forward Deployed Engineering, FDE）。这些工程师深入客户现场或以深度虚拟协作的方式直接对接业务一线。过去往往需要耗时 6 个月才能走通的实施流程，现在被大幅压缩至 2 到 4 周之内即可实现清晰的投资回报率（ROI）。
```mermaid
flowchart LR
accTitle: 前线部署工程落地流程
subgraph S1["现场调研与模式构建"]
A["现场/远程驻场"] --> B["解析运行形态<br/>Pattern of Life"]
end
subgraph S2["技术套件快速匹配"]
B --> C["定义客户核心痛点"]
C --> D["调用核心 API / SDK"]
end
subgraph S3["敏捷交付与验证"]
D --> E["4周内上线实时看板"]
end
```
这种速度的跃升源于工作范式的转变：FDE 工程师进驻后，首先倾听客户需求并基于物理空间建立“运行形态”（Pattern of Life）——即在数据层面厘清“谁在何时处于何地、什么物品在何时处于何处”。一旦客户明确了具体要解决的业务问题，工程师便能直接调用成熟的底层技术架构、API 以及 SDK 等工具链，迅速搭建定制化方案并即时呈现效果。以全球最大的发动机制造商之一为例，Worlds 团队仅用 4 周时间就完成了从零到全厂健康与安全实时监测看板的上线，能够对厂区内发生的异常与违规操作进行毫秒级实时捕捉与告警。
- **快速交付周期**：项目部署从以往的 6 个月缩减至 2 至 4 周。
- **超高精度模型验证**：在赌场合作伙伴的严苛场景中，AI 模型能够精准追踪与分析牌桌上各类筹码与游戏动作，量化展现模型执行具体任务的高效性。
---
### [16:02](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=962s) 构建传感器上下文：终结工业场景的海量误报
传统物理 AI 落地中最令企业头痛的问题之一就是海量的“误报”（False Positives）。许多企业受限于单一传感器的数据输入，例如单纯依赖仓库门禁传感器，导致管理系统一个月内可能收到多达 10 万次误报。Dave Copps 指出，消除这种警报疲劳的关键在于“通过传感器组合构建上下文（Forming context with sensors）”。
```mermaid
flowchart TD
accTitle: 多传感器上下文判别逻辑
A["门禁传感器: 检测到门处于开启状态"] --> C{"摄像头视觉判定"}
C -->|"泊位有货车正在作业"| D["状态: 占用中 (Occupied)<br/>结果: 正常装卸，不发警报"]
C -->|"货车已驶离/无车辆"| E["状态: 未占用 (Unoccupied)<br/>结果: 触发安全隐患警报"]
```
通过将视觉摄像头与门禁传感器进行数据联动，系统可以实现高级情境感知：当门禁传感器显示大门敞开，但摄像头检测到货位上正停着装卸货车时，系统判定为“大门开启且处于作业占用状态（Open and occupied）”，此时无需触发安全警报；而一旦货车驶离、泊位变空但大门依然敞开，系统则立即识别为“开启但无人看管（Open and unoccupied）”的安全隐患并鸣响警报。通过这种多源传感器的交叉验证，现场误报率能够直接“降到地板上”，彻底根除误报干扰。
---
### [16:48](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1008s) 突破传统视觉局限：无人机巡检与农业异常检测
在探讨无人机巡检与农业应用时，团队指出了传统计算机视觉（Legacy Computer Vision）的根本缺陷——“模型只能看到你训练它去看的东西”。Dave 打了一个生动的比方：如果一套系统只针对人行横道上的儿童进行了训练，那么当一头大象走过斑马线时，传统 AI 根本无法察觉大象的存在。而现代物理 AI 已经实现了范式转变，转向基于“异常检测”（Aberration Detection）的智能机制。
在智慧农业或管道检测中，系统无需穷举所有可能的损坏类型，而是通过学习健康作物的形态（例如健康苹果的表征）来运行。当果实出现褐变、病斑或结构异常时，系统能敏锐捕捉偏离正常基准的异常点。与此同时，物理 AI 还能将无人机巡检画面、天气数据、水质指标以及日常田间采集的多维数据在时间和空间维度上进行完全同步，实现真正的“传感器融合”（True Sensor Fusion），从而提供更深层次的农作物生长洞察。
---
### [18:16](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1096s) 时空同步的传感器融合与自主“运行形态”理解
虽然摄像头凭借既能识别物体又能判定“状态”（例如容器是满是空）的强大能力，依然是 Worlds 最青睐的核心传感器，但物理 AI 的终极形态在于跨模态的全传感器融合（Full Sensor Fusion）。在各类严苛甚至海外部署的环境中，系统将车辆车载 GPS、作业人员定位追踪器与全方位摄像头阵列无缝集成，统一纳入同一套 AI 系统进行全域感知。
物理 AI 的核心优势在于它对物理规律和运动模式的自主理解，这与传统计算机视觉对静态画面的浅层处理存在本质区别。系统能够自主学习物体在真实时空中的运动规律：例如当一架飞机着陆并缓缓滑入指定登机口时，系统会自动将这一系列连续的时空位移聚合识别为一次“进港事件”（Arrival Event）。企业无需预先编写繁琐僵化的规则代码去逐一限定，AI 便能自动提取规律并完成知识内化，为物理世界的复杂运营提供精准、自适应的智能化支持。
## Notable quotes
> [01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) the UI is a pane of glass — Dave Copps
> [01:09](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=69s) UI 就是现实世界本身 — Dave Copps
> [04:45](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=285s) 谁在何时出现在何地，物在何时处于何处 — Dave Copps
> [04:45](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=285s) 查找类似项 — Dave Copps
> [08:13](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=493s) AI 永远不会感到疲倦，它不知道什么是疲劳 — Chris
> [16:02](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=962s) 通过传感器组合构建上下文（Forming context with sensors） — Dave Copps
> [16:02](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=962s) 大门开启且处于作业占用状态（Open and occupied） — Dave Copps
> [16:02](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=962s) 开启但无人看管（Open and unoccupied） — Dave Copps
> [16:48](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1008s) 模型只能看到你训练它去看的东西 — Dave Copps
> [18:16](https://www.youtube.com/watch?v=RyuzJ9owYQk&t=1096s) 进港事件 — Dave Copps
## Mentioned
Dave Copps, Chris, Worlds, Ai4, 黄仁勋 (Jensen Huang), 英伟达 (NVIDIA), NVIDIA GTC, 埃森哲 (Accenture), Scale.com, Ken Spangler, 联邦快递 (FedEx), 阿姆斯特丹史基浦机场 (Schiphol Airport), 物理 AI (Physical AI), 大世界模型 (Large World Models), 生活模式 (Pattern of Life), 前线部署工程团队 (Forward Deployed Engineering, FDE)
## Video description
Dave Copps, CEO & Co-founder, Worlds
Abstracts
Join us to explore how Physical AI is transforming the daily operations of enterprises worldwide, helping them move beyond AI buzz to delivering real-time insights and results that drive increased productivity and enhance safety, while generating measurable ROI at scale.
[Source](https://www.youtube.com/watch?v=RyuzJ9owYQk)
Source quotes
[1:09] Dave Copps: the UI is a pane of glass
[1:09] Dave Copps: UI 就是现实世界本身
[4:45] Dave Copps: 谁在何时出现在何地，物在何时处于何处
[4:45] Dave Copps: 查找类似项
[8:13] Chris: AI 永远不会感到疲倦，它不知道什么是疲劳
[16:02] Dave Copps: 通过传感器组合构建上下文（Forming context with sensors）
[16:02] Dave Copps: 大门开启且处于作业占用状态（Open and occupied）
[16:02] Dave Copps: 开启但无人看管（Open and unoccupied）
[16:48] Dave Copps: 模型只能看到你训练它去看的东西
[18:16] Dave Copps: 进港事件
Source quotes
[1:09] Jensen Huang: it is time for AI to understand the laws of physics and predict physical occurrences
[1:09] Julie Sweet: the year of physical AI
[4:45] Dave Copps: Pattern of Life
[5:21] more like this
[14:13] What problem are you trying to solve?
[15:34] open and occupied
[15:34] open and unoccupied
[16:51] Dave Copps: It only sees what you train it to see.
[18:16] arrival event
Original transcript
[0:04] coming after the last uh speaker. It's
[0:05] actually pretty interesting because he's
[0:06] talking about how most our most pro AI
[0:09] projects are failing. So, this is a
[0:11] great intro into our our session. I
[0:12] think we've kind of latched on to uh a
[0:15] way that AI is not failing and it's
[0:17] proving ROI and it's proving uh to be
[0:18] very valuable and tangible ROI. So,
[0:20] we'll talk about that. Um great. So, um,
[0:24] yeah, I think in the past couple years,
[0:26] you know, it's been really hard to prove
[0:27] ROI, especially with Gen AI because it's
[0:29] it's kind of fuzzy and intangible,
[0:31] right? So, these, uh, I think, you know,
[0:33] most results are produced are very
[0:34] subjective and and I think, um, it's
[0:37] abstract like we feel like we're doing
[0:39] more productive. We feel like it's
[0:40] getting better, but we don't understand
[0:42] just how or why or what the value of
[0:43] that is. And so, now we have this idea
[0:45] of physical AI. So, bringing AI out of
[0:47] the back office into the real world and
[0:49] um, the ROI becomes more tangible
[0:51] because you can actually see the
[0:52] results. You can see and understand
[0:54] exactly the effect that things are
[0:55] having. You can do AB testing. Here's my
[0:57] environment before physical AI and
[0:59] here's my environment after physical AI.
[1:01] And you can see the difference. You can
[1:02] see the the uh improvement in
[1:03] productivity, efficiency, uh net safety,
[1:06] things like that. Um so let's back up a
[1:09] little bit and talk about just what is
[1:10] physical AI. you know, in in the broader
[1:12] sense, there's things like autonomous
[1:14] vehicles and robotics like a human
[1:16] robots are part of it, but really for
[1:17] the net for the for the most part, what
[1:19] we're going to be talking about today uh
[1:21] is is really just uh intelligent
[1:23] environments. So, it's it's it's our
[1:25] businesses. How do we create smart
[1:26] environments uh for our people and
[1:28] augment our people? So, physical AI,
[1:31] it's finally taking AI out of the back
[1:32] office and putting it directly into the
[1:34] physical operations of the world's
[1:36] largest companies. And so giving them
[1:37] the ability to measure and analyze and
[1:40] build AI automation directly into their
[1:42] physical environments, you know, so um
[1:45] the results of doing this, we start to
[1:47] now measure the real world. Now that we
[1:49] have the real world as data, we can
[1:50] start to see what's working, what's not
[1:52] working, and uh we can start to measure
[1:54] that and prove and see the tangible
[1:55] results that we're producing. So unlike
[1:58] back office AI where really kind of the
[1:59] the UI the UI is a pain of glass, the uh
[2:03] UI for physical AI is the real world.
[2:05] It's AI in the physical world and uh you
[2:08] know it occurs in in things like buzzing
[2:10] phones and and lights going off and and
[2:12] environments just working you know and
[2:15] um so it's a very very different thing.
[2:17] Um and I think the firing gun went off
[2:19] March this year. So uh Jensen Hang our
[2:23] rock star in AI uh they had the uh
[2:26] Nvidia conference and GTC conference in
[2:28] March uh in San Jose. we were there and
[2:31] uh he kind of fired the gun uh at the
[2:33] end of his uh keynote two-hour keynote
[2:35] and about 20,000 people in arena right
[2:38] uh he had a uh he basically had the
[2:40] slide that we all see where it's the
[2:41] lower left to upper right right on where
[2:43] AI has been where it's going and uh the
[2:46] upper right was physical AI and he said
[2:48] it's finally time for us to start to
[2:50] interact with AI that understand the
[2:51] laws of physics and help us to predict
[2:54] to not only understand what's happening
[2:56] in our physical environments but even to
[2:57] start to predict what might
[2:59] Um he was followed very quickly by
[3:01] Julie. He's the CEO of Accenture and she
[3:04] came out and said this is the year of
[3:05] physical AI, one of our partners. Uh and
[3:07] so we're excited about that working with
[3:09] Accenture. Um but I think you know this
[3:11] the timing of this is perfect, right?
[3:13] Because really when you think large
[3:14] language models, we were talking about
[3:15] that in the previous session, right?
[3:17] Language models are great, but we've
[3:18] kind of we've run out of data. I mean
[3:21] for the for for the most part uh you
[3:23] know we've uh put given the language
[3:25] models all the data in the world. uh and
[3:27] we've and that's it, you know, but now
[3:30] we'll make more data, but for all
[3:31] intents and purposes, we've run out of
[3:33] training data. So, it just turns out
[3:35] that there's more training data locked
[3:37] in the physical world than there will
[3:38] ever be in documents. So, we kind of see
[3:41] the work we're doing uh with physical AI
[3:43] as a real unlock for the next wave of AI
[3:46] and helping AI to understand our world.
[3:48] It turns out the best way to learn about
[3:50] the world is not just by reading about
[3:51] it. It's about giving it's it's by
[3:53] experiencing it like we do. So giving AI
[3:56] by connecting into existing cameras and
[3:57] sensors, we give AI the ability to start
[3:59] to experience the world like we do and
[4:01] learn in a way that it's never learned
[4:02] before.
[4:04] Um,
[4:07] more. Yeah. And this is so uh the these
[4:10] are headlines I pulled just a few days
[4:12] ago, but they're all in the last eight
[4:14] weeks. So you can start to see the the
[4:16] momentum's building. It's all the way
[4:17] from, you know, Aentic AI and physical
[4:19] AI being a multi-t trillion dollar
[4:20] business to how physical AI is going to
[4:22] reshape the world to now stocks. They're
[4:24] even starting to talk about physical AI
[4:26] and stocks. Of course, you know, is is
[4:28] is physical AI going to finally be the
[4:30] way that we bring value uh into the
[4:32] world. And so, uh it's definitely
[4:34] hitting right now. Um I kind of
[4:36] predicted last year I thought it would
[4:37] happen in the first quarter, you know, I
[4:39] think it's really happening right now.
[4:40] Like as we're sitting here now that the
[4:42] momentum's picked up hugely.
[4:45] Um so the way we are doing it is we're
[4:47] actually building large world models
[4:48] instead of large language models. We're
[4:49] actually connecting our AI platform
[4:52] directly into an organization's existing
[4:53] cameras and sensors and ex taking what's
[4:56] being captured by those those sensors
[4:58] and and converting it into a live data
[5:00] stream. So start to think about it. If
[5:02] you can actually have all the activities
[5:04] that are driving value in your business
[5:06] uh converted into live data, now you can
[5:08] start to build the applications and
[5:09] agents on top of the platform to make uh
[5:12] magic happen. Um so one of the big
[5:14] learnings that we had kind of going to
[5:17] market and I think there's a new
[5:18] baseline for physical AI when we walk
[5:20] into any any industry it doesn't even
[5:23] matter who the client is we always start
[5:26] with pattern of life. So what I mean by
[5:28] that which stole us from the military by
[5:30] the way uh the military kind of came up
[5:32] with this idea of if we can start to
[5:34] understand who's where when and what's
[5:36] where when they call that pattern of
[5:38] life. Well so we started with that every
[5:40] every engagement that we have is the
[5:42] first thing we want to do is connect
[5:43] into your existing cameras and sensors
[5:44] and start to establish pattern of life.
[5:46] Who's where when's where when. If we
[5:48] have that as data now the plus comes in.
[5:51] Now we can start to build the
[5:52] applications and agents on top of that
[5:53] data to solve real problems. Um so the
[5:56] result is you know the new uh live
[5:59] spatial data when I say live spatial
[6:00] data I mean what's being sensed when
[6:03] it's being sensed and where it's being
[6:05] sensed. So live spatial data it's a new
[6:07] raw material for digital and then uh
[6:10] once we have that data then a AI becomes
[6:12] the transformation engine the
[6:13] intelligence engine that we use to start
[6:15] to build the agents and applications
[6:17] that will solve solve an organization's
[6:18] problems. Um, so setting the stage for
[6:22] physical AI, you have to do a few
[6:24] things. And so we kind of learned the
[6:25] lesson the hard way. You can't just do
[6:27] one of these. You've got to really do
[6:28] them all. And so our platform is really
[6:30] probably the first platform that is the
[6:33] an end to end AI platform for physical
[6:35] AI. So we start with training. So we've
[6:37] actually got the ability. So you
[6:39] probably seen how you train AI is
[6:41] typically you have people drawing boxes.
[6:43] This is a dog. This is a cat. This is a
[6:45] forklift, you know, and you it takes
[6:47] hours and hours and hours. Companies
[6:48] like uh scale.com made billions of
[6:51] dollars with hundreds of thousands of
[6:53] people around the world drawing boxes
[6:54] teaching AI about the world. Well, we
[6:56] decided that was too cumbersome. And so
[6:58] we flipped the entire thing on its head.
[7:00] So we built a system that can train AIs
[7:02] in literally minutes, minutes to hours.
[7:05] And so it's a from a instead of a top-
[7:07] down manual human and loop process, it's
[7:09] now a bottoms up fully automated AI
[7:11] system. You turn on your cameras, it
[7:13] starts to index everything it sees
[7:15] automatically with no human
[7:16] intervention. and then gives you kind of
[7:17] a search and clustering tool that lets
[7:19] you kind of choose the things that are
[7:20] most important to you and very quickly
[7:22] annotate your world. So teaching AI
[7:24] about your world, we can do it in the
[7:25] same day. We can go into a client's site
[7:27] and teach our AI about the world in the
[7:29] same day. And of course, inference is
[7:31] just taking those models now and
[7:32] applying it towards the the data that's
[7:34] being captured by all the sensors and
[7:36] turning that into into live spatial
[7:38] data. And again, once you have that, now
[7:40] it's just about building the apps and
[7:42] agents and agentic architectures on top
[7:44] of that data. The newer stuff we're
[7:46] doing that's showing serious ROI is um
[7:49] is a Gentic AI. So we're we're now
[7:52] taking the once we establish pattern of
[7:53] life and that base data, we're now have
[7:56] an FTE team, Ford engineering team that
[7:58] can go on site and very very quickly uh
[8:00] start to within a matter of two or three
[8:03] weeks start to show uh very quick time
[8:05] to value and return on investment. So at
[8:08] this point uh I'm going to turn it over
[8:10] to Chris and show you some of the stuff
[8:12] we're working on.
[8:13] >> Cool. Thank Thank you.
[8:15] um we thought we would rapid fire
[8:17] through just a few examples um you know
[8:20] specifically around RARI and RARI begins
[8:22] with actually training the AI agents
[8:24] themselves and so this specific example
[8:26] is uh is inspired by our friend Ken
[8:28] Spangler he was CIO of FedEx uh before
[8:31] he recently retired and while we can't
[8:32] show you FedEx we'll show you the Chipel
[8:34] airport and so we start with just raw
[8:37] video and so the first piece is we just
[8:38] load raw video in the system inq
[8:41] automatically identifies all the objects
[8:43] and all the video and then the user
[8:45] selects what's important to them. So
[8:46] they select the aircraft and they click
[8:48] on more like this and it acts like a
[8:50] magnet to attract all the related
[8:52] imagery. So you end up with all the
[8:54] things that that look like aircraft that
[8:55] are similar shape, color, size, texture,
[8:58] even things that move in a similar
[9:00] manner. And that's what you're seeing
[9:01] here. And so literally um we do this
[9:03] demonstration live on stage for clients
[9:05] data regularly. And literally in a
[9:07] couple minutes we have 1300 examples.
[9:09] And so that's pretty incredible given
[9:11] the way people normally train AIs. I
[9:14] mean, normally this would take hundreds
[9:15] of hours worth of effort. And so, um,
[9:18] instant ROI up front. But if we talk
[9:20] about how this then applies to our
[9:22] clients, um, this is this is one
[9:24] example. So, a security profile where,
[9:27] you know, a single individual can't
[9:29] watch hundreds of cameras. They just
[9:30] can't do it. And the problem doesn't get
[9:32] any better if you put a room full of
[9:33] people watching hundreds of cameras
[9:35] here. And so really what you need is you
[9:38] need AI watching the cameras. And so
[9:41] there's a lot of really interesting
[9:42] things that you can observe. I mean
[9:43] whether it's an entry event, an exit
[9:44] event, a perimeter breach, people
[9:46] jumping over fences. Dave mentioned the
[9:48] the anomaly detection is really the most
[9:52] interesting thing. So the AI itself can
[9:54] establish a pattern of life, understand
[9:56] what's normal, what's not normal, and
[9:57] then alert you just of the things that
[9:59] are abnormal. And so maybe the best
[10:01] part. And so and the final piece is AIS
[10:03] don't get tired. um they don't
[10:05] experience fatigue and it's 110 degrees
[10:07] outside and you have you have uh
[10:08] security folks outside and you know
[10:11] again that's that's a tough job and so
[10:12] this is an example of a security example
[10:15] in a hospital and so if we zoom in on
[10:17] the hospital uh we can actually you know
[10:19] access the system here uh and you get a
[10:22] you get an overview and so in this case
[10:23] this is the floor of a hospital uh we
[10:25] can turn on things like like the floor
[10:27] plan itself you see all the cameras on
[10:29] the floor uh we have the the span of
[10:31] view of the cameras that you can see
[10:33] here and then you have the different AIs
[10:34] that are deployed. And so we've trained
[10:35] it. In this case, we look for doctors or
[10:37] nurses or patients or visitors. And
[10:40] that's kind of what you're seeing here.
[10:41] This is our our our application. But the
[10:43] whole point of this is not to actually
[10:45] watch the video. What you want to do is
[10:47] have the AI watch the video, interpret
[10:50] that, and then ultimately put it back
[10:51] into view here. And so in this case,
[10:53] this is what an analyst would actually
[10:54] see. So the blue dots are people moving
[10:57] around, a combination of of patients and
[10:59] visitors. In this case, we're featuring
[11:01] we're color coding the security
[11:02] professionals. And you can see that in
[11:04] the right, you know, the in in the red
[11:06] there. And so, um, super effective. And
[11:08] you look at the I mean, the return on
[11:09] investment is really significant. The
[11:11] ROI, um, you know, it really speaks for
[11:15] itself. You have a kind of a 10x
[11:16] improvement in terms of the ability for
[11:18] people to do capacity, in terms of to do
[11:20] their work. And then the final piece is
[11:22] the alarm reduction. The the false
[11:24] alarms that you get in these security
[11:27] systems dramatically reduces. And we've
[11:29] seen this 80 or 90% which is really
[11:31] incredible. And that's additional
[11:32] savings there. And so this next example
[11:34] is um we have a top five trucking
[11:36] company that's one of our clients uh in
[11:38] logistics. We have uh we're automating
[11:41] their terminals around the country right
[11:42] now. So 18 terminals around the country.
[11:44] We have AI deployed at the Guardgate um
[11:47] and you know we're moving into yard
[11:48] management and then ultimately in the
[11:49] maintenance facilities. But you can see
[11:51] here we um AIS are detecting the the the
[11:55] unique signatures on each truck. We're
[11:57] doing the same thing with the container
[11:58] IDs here and reading them off the sides
[11:59] of the containers. We're authenticating
[12:01] the logos and the material on the the
[12:03] sides of the u the payloads as well. And
[12:05] so, you know, if all three things match,
[12:07] we let them flow right through. And
[12:08] that's what you're seeing here. The uh
[12:10] increase in productivity here is, you
[12:12] know, 92% which is really significant.
[12:14] So, they're not sit down on the side of
[12:15] the road. And of course, you end up, you
[12:17] know, going from 7 minutes to 10 seconds
[12:19] of check-in process, which is which is
[12:21] awesome. Maybe the biggest thing is
[12:23] increased uh driver satisfaction. I
[12:26] mean, ultimately, these guys are they're
[12:28] they're paid by the mile. They're not
[12:30] paid by the hour. So you see them
[12:31] stacked up outside these big big uh big
[12:34] facilities. And so that's completely
[12:36] made. And then the final piece is that
[12:37] they're not tugging the wrong payload
[12:39] and figuring that out halfway around the
[12:41] which is obviously important. Um you
[12:44] this is a big steel manufacturer. So a
[12:46] top five steel manufacturer in the
[12:47] country. Uh we have three different use
[12:49] cases. We regularly see that from
[12:51] clients. We'll have one use case that'll
[12:52] have the biggest ROI and then we have
[12:53] some tagalong use cases. The first is u
[12:56] we're making more accurate cuts and prec
[12:58] precise cuts in terms of the steel and
[13:00] so they're getting better yield out of
[13:01] the steel. We're making sure that the
[13:03] that the steel itself is straight and so
[13:05] uh being able to detect defects very
[13:07] early in the process. The second use
[13:08] case is a safety use case. This big
[13:10] cauldron we're using a thermal camera to
[13:12] see right through that cauldron and
[13:14] making sure that it's not overfilled.
[13:15] And the final use case is we're keeping
[13:17] the the worker out of harm's way. And so
[13:19] we see that down below. And so, you
[13:21] know, the use case that pays for it is,
[13:23] you know, we detect a defect very early
[13:25] in the process. Things aren't straight.
[13:26] They aren't exactly what they are. We
[13:28] save $2,300 every time we every time we
[13:30] detect that. And so, that's phenomenal.
[13:32] We love that. And then the final use
[13:34] case is um this is one of the largest
[13:36] pipeline providers in the country. And
[13:38] so, we're doing leak detection. Um and
[13:40] this is interesting. What you see here
[13:41] is a thermal camera detecting even very
[13:44] very small leaks. Uh and the thermal
[13:46] camera is the red. And then we have
[13:47] change detection down below. We find
[13:49] that the combination the algorithms
[13:50] together really provides a really nice I
[13:54] mean it detects leaks very very early on
[13:55] and we can remediate those and so you
[13:57] know even uh even very small leaks left
[14:00] over prolonged period of time can result
[14:02] in you know literally over a million
[14:04] dollars worth of worth of investment. So
[14:06] um we'll we'll stop there. That was kind
[14:09] of rapid fire.
[14:10] >> So at this point if you have any
[14:11] questions we'll just uh take your
[14:13] questions.
[14:14] >> Yeah for sure. Sure. And I I'll tell you
[14:16] what, if you asked that question just a
[14:17] year ago, we would give you a very
[14:18] different answer. Uh this year, um we in
[14:22] January, we decided to create a what we
[14:23] call forward deployed engineering team.
[14:25] Some of you may have heard of things
[14:26] like this, but it's a forward deployed
[14:27] engineers to actually go on either go on
[14:29] site or virtual do it virtually, but
[14:31] actually um work with the client to
[14:33] create very fast ROI. So things that
[14:35] were taking us literally six months last
[14:37] year like we're doing now in two to four
[14:40] weeks you know and the reason we're able
[14:41] to do that is because with these FTEEs
[14:44] essentially what we've done is um when
[14:46] they go on site it's listening to the
[14:48] customer what I said earlier about
[14:49] pattern of life we establish pattern of
[14:51] life so who's where when what's where
[14:52] when as data right and now our FDS asked
[14:55] the question to the customer what
[14:56] problem are you trying to solve right
[14:58] and so it's now it's a description of
[15:00] the problem and now it's what
[15:01] application now we have our core
[15:02] technology we have our API I we have our
[15:05] SDK all these things that are these
[15:06] tools at our disposal now to create what
[15:08] they want and to show the results
[15:10] immediately. Um we had a one of the
[15:12] largest engine manufacturers in the
[15:14] world as a client. We went from zero to
[15:17] uh having a live dashboard for health
[15:19] and safety on their entire plant showing
[15:22] exactly when the bad things were
[15:24] happening as they're happening in real
[15:25] time. We did that in four weeks. Last
[15:27] year if I'm just honest about it that
[15:29] would have taken six months. So we're
[15:31] able to show our I mean I don't know if
[15:33] you want to add to that but
[15:34] >> yeah and I think the quality spec
[15:36] specifically right you you have the
[15:37] quality of the AI itself of of the
[15:40] models and so we have a lot of a lot of
[15:42] ways that we can we can quantify and
[15:44] tell you exactly how how well the model
[15:46] is doing. In fact we have models that
[15:48] are running in casinos. So one of our
[15:50] partners we actually analyze gameplay
[15:51] and so we have to be very specific
[15:53] because we're analyzing all the chips
[15:54] and all the things. So you get very
[15:56] specific and analyze, you know, am I
[15:58] doing the thing that I that I said I was
[15:59] doing and how effective am I that? So
[16:02] >> sometimes we're in environments because
[16:03] like Chris talked about false positives
[16:05] earlier, you know, like that's always
[16:06] been a big thing, you know, with AI,
[16:08] especially physical AI. Um, a lot of
[16:10] it's because they're relying on one
[16:12] sensor, right? If it's a door sensor and
[16:14] it's whatever, but they we talked to
[16:16] people that have 100,000 false positives
[16:17] a month.
[16:18] >> And so we've come up with this idea of
[16:19] forming context with sensors. Like if
[16:21] you start to just add sensors together.
[16:23] So I have a camera that's looking at a
[16:24] door and I have a door sensor. If the
[16:26] door sensor says the door is open but
[16:27] there's a truck there. I don't need to
[16:29] send the alarm. But the moment the truck
[16:31] pulls away now it's a security concern.
[16:34] Right? So by combining the sensors
[16:35] together we can start to say the door is
[16:37] open and occupied. Don't send an alarm.
[16:39] When the truck leaves away it's open and
[16:40] unoccupied. Sound the alarm. You know so
[16:43] the false positives go to the floor.
[16:45] Just don't have them anymore.
[16:48] >> Anybody else?
[16:49] >> Yeah. Right here. Yeah.
[16:51] >> Yeah. Love that. And we doing some
[16:53] starting to do some drone work. So you
[16:55] have a drone fly over. We can do the
[16:57] analysis of, you know, what's going on
[16:59] there. And so we had a, you know, a
[17:01] pipeline piece that we're inspecting a
[17:02] pipeline and understanding in a similar
[17:04] way. But agriculture will be very
[17:05] interesting. Like we Yeah.
[17:06] >> You talked about earlier um you know,
[17:09] one of the kind of the faults of let's
[17:11] call it legacy computer vision like
[17:13] what's happened the last few years is it
[17:15] only sees what you train it to see,
[17:17] right? And so um you had a crosswalk and
[17:21] uh for k children in a crosswalk people
[17:23] and an elephant walked by the AI
[17:24] wouldn't see the elephant because it
[17:26] wasn't trained to see it you know. So um
[17:28] but like in the case of agriculture like
[17:30] one of the things is training a system
[17:32] on what a healthy fruit or a healthy
[17:34] apple looks like and just say let me
[17:36] know when something doesn't look like
[17:38] that like it's an aberration detection
[17:40] you know um or give it lots of examples
[17:42] of browning or lots of examples of this.
[17:44] And so I think it's a different world
[17:45] today. the AIS are now able to do much
[17:47] much more than they could have just a
[17:48] few years ago. Um, and then also I think
[17:50] the what you said earlier like com
[17:52] combining the things together now like
[17:54] now I can take weather and I can take
[17:55] water and I can take you know data I'm
[17:57] collecting every day in the field and uh
[17:59] and I can start to make sense of that.
[18:01] This idea of kind of a um true sensor
[18:05] fusion like we can literally take all
[18:06] that data pull together in space and
[18:08] time so everything's synchronized in
[18:10] space and time so you can start to have
[18:11] real intelligence about crops and uh
[18:14] things like that. But yeah,
[18:15] >> interesting.
[18:16] >> Yeah, cameras are our favorite sensor
[18:17] because they're very powerful, right?
[18:19] Powerful. You can not just identify
[18:21] things, but you can identify state like
[18:23] something's full or empty or you know,
[18:25] things like that. So, cameras are our
[18:26] favorite sensor, but no, we actually are
[18:28] uh we can execute a full sensor fusion.
[18:30] You want to talk a little bit about the
[18:31] work we're doing in overseas?
[18:34] >> Yeah, I mean, we incorporate a wide
[18:35] number of sensors. So, we have things
[18:36] like like vehicle GPS sensors on
[18:39] vehicles. We have personnel trackers on
[18:41] personnel, right, on on people. we can
[18:43] identify with cameras as well. So really
[18:46] understanding the world uh through the
[18:49] lens of a bunch of different sensors and
[18:51] but applying a single set of AIs on top
[18:52] of that is really what we're what we're
[18:54] after. And certainly certainly the
[18:55] physics and the way things move is is
[18:58] much different than just kind of a a
[19:00] general um kind of classical computer
[19:03] vision piece, right? Yeah.
[19:05] >> Yeah. And we learn things like pattern
[19:07] of life and how how things move and the
[19:09] pattern in which it moves. You have an
[19:10] aircraft that lands and pulls in a gate
[19:12] and that's an arrival event, right? You
[19:14] don't have to tell it that. I mean, you
[19:16] do it would say these hundred things are
[19:18] all the same thing and you would go
[19:19] that's an arrival event and then you
[19:21] would learn that, right? So, um yeah.
[19:24] Yeah.
[19:26] Okay. Cool. Thank you very much.
[19:27] Appreciate it. Thanks.
