Physical AI
AI that operates in the real world. Robots that walk, vehicles that drive themselves, drones that fly through contested airspace, surgical systems that work to sub-millimeter precision. Start with the landscape, or jump into the layer that interests you: governance, hardware, software, systems.
The overview
Read this first. The full map of Physical AI in one document.
The four layer guides
Each layer guide is a self-contained deep dive. Read them in any order; all cross-reference each other.
Layer 1: Governance
ReferenceThe rules: who controls the robot, how safety is ensured, who is liable. The EU AI Act and the May 2026 Omnibus simplification, the Machinery Regulation, the Revised Product Liability Directive, NIST AI RMF and the AI Agent Standards Initiative, ISO 10218:2025, NHTSA FMVSS modernization, the SELF DRIVE Act, FDA pathways, NVIDIA NemoClaw, six governance gaps, the Intent Layer thesis, and the regulatory timeline through 2028.
Layer 2: Hardware
ReferenceThe body: sensors (RGB, depth, LIDAR, audio, IMU, force), compute (the NVIDIA Jetson ladder from Orin Nano Super at 67 TOPS for $249 to Thor at 2,070 FP4 TFLOPS for humanoid-grade foundation models), actuators (servos, DC, stepper, grippers, rotors), and structure (PLA, PETG, resin, aluminum extrusion, carbon fiber). Verified products and prices.
Layer 3: Software
ReferenceThe brain: runtime (JetPack 6.2, CUDA 12.6, TensorRT 10.3), perception models (YOLO26, Whisper Turbo, DeepStream, SAM3), decision and action models (GR00T N1.7, imitation vs reinforcement learning, Nav2, MoveIt 2), world models (Cosmos 3.0 vs V-JEPA 2.1), and frameworks (ROS2, the Isaac platform, Newton physics, LeRobot, AgenticROS, NemoClaw).
Layer 4: Systems
ReferenceWhat gets built: robotic arms (SO-ARM101 to FANUC CRX, Universal Robots, ABB), mobile robots (AMRs, RaaS), autonomous vehicles (Tesla FSD 14.3 and Cybercab production, Waymo, Zoox, Aurora), humanoid robots (Unitree G1 at $16K and H2 shipping April 2026, Figure 03, Agility Digit, Boston Dynamics Atlas, 1X NEO, Tesla Optimus), drones (Skydio, Anduril, DJI), and surgical / medical robots (Intuitive da Vinci, Medtronic Hugo, CMR Versius, Stryker Mako).
The personal plans
A practical pair for anyone serious about getting hands-on. Read first, then sequence the hardware.
Physical AI Learning Plan
PlanA read-first four-phase study plan moving from foundations (the landscape, the NVIDIA "three computers" concept) through the software stack (Jetson AI Lab, Isaac Sim / Lab / ROS / GR00T, Hugging Face LeRobot, ROS2, LeCun JEPA, AgenticROS) and hardware (Jetson Orin Nano Super, SO-ARM101, Reachy Mini) to first builds. Each phase has a clear exit criterion. Spending time in Phase 1 and 2 before touching hardware is the right move.
Physical AI Build Plan
PlanA sequenced hardware purchase + project plan. Three hardware tiers from ~$1,000 starter (Jetson Orin Nano Super + SO-ARM101 + Reachy Mini Wireless) through expansion (depth camera, multi-camera vision, voice assistant, JetBot) to aspirational humanoid-grade. Six projects build on each other: Jetson hello-world, robotic-arm imitation learning, AgenticROS Claude-to-robot, multi-camera vision, offline voice assistant, autonomous navigation, sim-to-real in Isaac Sim.