djvu Physical AI: Engineering Embodied Intelligence for the Real World

Digital AI gets to skip physics. Physical AI cannot.

Every grasp, every step, every navigation decision in the real world is shaped by gravity, friction, latency, sensor noise, and battery limits — constraints that no amount of model scale can paper over. Physical AI is the engineering book for the people building the systems that act in the world.

It walks the full stack between "the model understands the word 'grasp'" and "the robot picked up the cup without crushing it." Token-level reasoning meeting joint torques. Vision-language-action models meeting Jetson thermal budgets. Sim-to-real bridges, data flywheels, and the regulation that's catching up.

• The grounding problem in physical form — why an LLM's statistical "grasp" is not yet a motor command
• Five phases of robotic foundation models, from CLIP bolt-ons to GR00T and Helix
• The Physical AI stack: sensors, perception, world models, policy, actuators at 10–50 Hz
• Architectures: VFMs (DINOv2, SAM), VLMs (CLIP, PaliGemma), VLAs (RT-2, OpenVLA, Pi-0)
• Diffusion policies and Liquid Neural Networks — when each fits
• The 1,000× cost of physical training data and the teleoperation rigs (Aloha, UMI, GELLO) that scale it
• Open X-Embodiment as robotics' ImageNet moment
• Bridging sim-to-real with NVIDIA Isaac Sim, MuJoCo, domain randomization, and NVIDIA Cosmos world models
• Edge compute realities: Jetson Orin, Thor, Hailo, Qualcomm RB6 — and the quantization that fits a 7B VLA into a 60W envelope
• Safety and regulation in 2026: EU AI Act, Colorado AI Act, ISO 10218, ISO/TS 15066, ISO 26262, SAE J3016
• Industry deployments: Covariant, Symbotic, Boston Dynamics Stretch, Waymo, Wayve, Tesla FSD, John Deere autonomy
• The frontier humanoids: Tesla Optimus, Figure, Atlas, Apptronik Apollo, Agility Digit, Unitree H1/G1, 1X Neo, Sanctuary Phoenix
• Goldman Sachs market projections, geopolitical asymmetry, and the scaling-law question for embodied AI

Includes four reference appendices: a glossary bridging robotics and ML vocabulary, an annotated foundational reading list, a hardware buyer's guide with 2026 prices, and a directory of open-source models, simulators, and datasets.

Code examples in Python, YAML, and C++. Nearly 480 pages of citations and real engineering — every named model, dataset, paper, and regulation backed by an inline citation and a per-chapter References section.

Written for robotics engineers transitioning from classical pipelines, ML engineers from the LLM world bridging into embodied systems, and infrastructure architects evaluating where Physical AI fits.

A companion to the LLM Primer series.