- Watches a video of a task being performed (by a human or another robot).
- Understands the task using vision-language model capabilities.
- Plans execution by reasoning about object affordances, physics, and goal states.
- Executes the task on a real robot, adapting to its specific embodiment.
Traditional robot training requires 1,000+ hours of robot teleoperation data per task. Skild's approach: 1 video, infinite robots. The implications are significant for:
- Industrial robots: factory robots learning new assembly tasks in minutes instead of weeks.
- Household robots: home robots learning new chores from YouTube videos.
- Surgical robots: surgical assistants learning new procedures from surgeon demonstrations.
- Agricultural robots: harvesting robots learning new crop varieties from farmer demonstrations.
- Last-mile delivery: delivery robots adapting to new environments from video.
Jetson Thor robotics computer
Jetson Thor is the next-generation robotics computer, announced in 2024 and shipping in volume in 2025-2026:
| Specification | Jetson Thor | Jetson Orin (prior gen) |
|---|
| GPU | NVIDIA Blackwell | NVIDIA Ampere |
| AI performance | 2,000 TOPS | 275 TOPS (Orin AGX) |
| CPU | 14-core Arm Cortex-A78AE | 12-core Arm Cortex-A78AE |
| Memory | 64GB LPDDR5X | 32GB LPDDR5 |
| Power | 40-130W | 15-60W |
| Compute increase | — | ~7x more AI performance |
Source: NVIDIA Jetson product specifications (2026).
Jetson Thor target applications
- Humanoid robots: Figure 02, 1X Neo, Sanctuary Phoenix, Apptronik Apollo, Tesla Optimus, Agility Digit, Unitree H1.
- Autonomous Mobile Robots (AMRs): warehouse robots, hospital delivery, hotel service.
- Industrial automation: assembly, welding, painting, inspection.
- Surgical robots: Intuitive Surgical da Vinci, Medtronic Hugo, Johnson & Johnson Ottava.
- Agricultural robots: John Deere autonomous tractors, harvesters.
- Last-mile delivery: Nuro, Starship Technologies, Kiwibot.
- Defense and security: Boston Dynamics Spot, Ghost Robotics Vision 60.
- Service robots: restaurant, retail, hospitality robots.
Jetson Thor pricing: $3,000-$5,000 per module (vs $2,000-$4,000 for Jetson Orin). Volume discounts available for OEMs (NVIDIA, 2026).
Cosmos world foundation model
Cosmos is NVIDIA's world foundation model (WFM) for physical AI. WFMs generate physics-based video simulations of environments, enabling robot training in simulated worlds without the cost and risk of real-world training.
Cosmos key features
- Physics accuracy: models gravity, friction, collisions, deformable objects, fluid dynamics.
- Sensor simulation: generates RGB cameras, depth, lidar, radar outputs.
- Domain randomization: varies lighting, weather, object positions, materials for robust training.
- Camera control: precise control of camera trajectory, focal length, and aperture.
- Long-horizon generation: generates extended video sequences for training.
- Integration: integrates with Isaac Sim, Isaac Lab, and DRIVE Sim for full-stack simulation.
Cosmos applications
- Robotaxi simulation: rare event simulation (children running into street, emergency vehicles, unusual weather).
- Humanoid robot training: manipulation tasks, bipedal locomotion, human-robot interaction.
- Industrial automation: assembly line variations, fault recovery, quality inspection.
- Autonomous trucking: long-tail scenarios, weather variations, construction zones.
- Surgical simulation: rare complication simulation, instrument handling.
Isaac Sim and Isaac Lab updates
NVIDIA Isaac Sim (Omniverse-based robotics simulation) and Isaac Lab (robot learning framework) were updated September 10, 2026:
Isaac Sim updates
- New sensor models: more accurate camera, lidar, IMU, force/torque sensor simulation.
- Improved physics: better soft body, fluid, and granular media simulation.
- Reinforcement learning support: tighter integration with RL training workflows.
- Synthetic data generation: scalable data generation with domain randomization.
Isaac Lab updates
- New robot models: support for the latest humanoid robots, AMRs, and industrial arms.
- Sim-to-real transfer: improved techniques for transferring trained policies to real robots.
- Imitation learning: tighter integration with teleoperation data collection.
- Multi-robot learning: support for coordinating multiple robots.
Physical AI market context
The physical AI market is rapidly growing:
- Total market: $50B+ in 2026; projected $300B+ by 2030 (Goldman Sachs, McKinsey).
- Robotaxi market: $5B+ in 2026; projected $200B+ by 2030.
- Humanoid robot market: $2B+ in 2026 (mostly R&D); projected $30B+ by 2030.
- Industrial automation: $30B+ in 2026; projected $80B+ by 2030.
- AMR/warehouse robot market: $8B+ in 2026; projected $30B+ by 2030.
NVIDIA is positioned as the primary compute and platform provider, with potential to capture 30-50% of the physical AI compute market. NVIDIA's automotive and robotics revenue: $2B+ in FY2026, projected $5-10B+ by FY2028 (NVIDIA; Goldman Sachs, 2026).
FAQ
Is robotaxi commercially available in 2026?
Limited commercial robotaxi deployments exist in 2026: Waymo in Phoenix, San Francisco, Los Angeles, Austin; Cruise (limited) in San Francisco; Pony.ai in Beijing, Shanghai, Guangzhou, Shenzhen; AutoX in Shenzhen, Beijing; WeRide in Beijing, Guangzhou. Most operate within specific geofenced areas with safety drivers available. Mass-market L4 robotaxi (no safety driver, any location) is projected for 2028-2030. DRIVE Hyperion 9 targets this 2027-2028 commercial deployment window (NVIDIA, 2026).
How does Skild AI's single-video training compare to other robot learning approaches?
Robot learning approaches comparison: (1) Skild AI single-video - 1 video, generalizes across tasks/robots. (2) Behavior cloning - thousands of teleoperation demonstrations per task. (3) Reinforcement learning - millions of simulation episodes per task. (4) Imitation learning from play - hours of human demonstration data. (5) Self-supervised learning - large amounts of unlabeled robot interaction data. Skild's approach is most efficient in terms of data requirements but requires powerful foundation models and may not generalize to all physical tasks. The single-video demo is a research milestone; production deployment is still 1-2 years away (Skild AI; NVIDIA, 2026).
Fazlur Rahman
