Everything about Physical AI

0. What Physical AI actually is

Physical AI is NVIDIA's umbrella term for AI that moves from the digital world into the physical world. Instead of chatbots generating text or code assistants writing software, Physical AI powers robots that walk, vehicles that drive themselves, drones that navigate contested airspace, surgical systems that operate with sub-millimeter precision, and factory floors that run autonomously.

Jensen Huang at GTC 2026 (March 16, 2026): "Physical AI has arrived. Every industrial company will become a robotics company."

The shift is fundamental. Software AI (LLMs, code assistants, image generators) operates in the digital domain: text in, text out. Physical AI operates in the real world: sensor data in (cameras, LIDAR, microphones, force sensors), motor commands out (wheels, joints, grippers, rotors). The AI must perceive, reason, plan, and act in real time, in environments it has never seen before, with consequences for failure that are physical: a robot drops a patient, a drone hits a building, an autonomous truck jackknifes.

This is why simulation, edge computing, and foundation models all converge in Physical AI. You cannot train a humanoid robot to walk by having it fall down 10,000 times in a real lab. You train it in simulation (Isaac Sim), using a physics engine (Newton), with synthetic data (Cosmos), then transfer the learned skills to real hardware (Jetson). The sim-to-real pipeline is the central engineering challenge.

Source: NVIDIA Robotics platform.

1. Key terms (quick reference)

These terms appear throughout this document. Understanding them upfront makes everything that follows click. A full glossary is at the end.

2. The NVIDIA ecosystem

NVIDIA has built the dominant end-to-end platform for Physical AI. Understanding this stack is essential because nearly every serious robotics company builds on some part of it.

The ecosystem is organized into four layers: solutions (what problem are you solving), software (how do you build it), hardware (what runs it), and developer resources (where do you learn).

Software layer (where the depth is)

The Isaac platform is a family of tools, each handling a different stage of the robot development lifecycle.

Developer layer (where to learn)

• Jetson AI Lab: Hands-on tutorials for GenAI on Jetson. LLMs, VLMs, VLAs, speech, applications. The #1 starting point.
• Jetson Developer Portal: Central hub for documentation, downloads, forums, ecosystem partners.
• NVIDIA Robotics Blog: Technical blog posts on robotics tools, SDKs, workflows.
• GR00T GitHub: Open-source foundation model code (N1.7 as of May 2026).

3. The hardware ladder: Jetson family

The Jetson lineup is a performance ladder. Each tier is a different price/performance point. JetPack SDK is unified across all of them: code written on one runs on any other.

Currently relevant (for new projects in 2026)

Industrial variants

IGX Platform: industrial-grade edge AI. Blackwell-based Thor systems with safety certifications for factories, medical devices, autonomous vehicles. Enterprise deployments.

Buy: Jetson Store | Orin Nano Super Dev Kit.

4. The software stack: Isaac platform

Isaac Sim (simulation)

The digital twin environment where robots learn before touching reality. Key capabilities:

• Photorealistic rendering: synthetic camera images indistinguishable from real photos. Robots trained on synthetic data transfer to real environments.
• Physics simulation: powered by PhysX and now Newton. Realistic gravity, friction, collisions, deformation.
• Sensor simulation: simulate cameras, LIDAR, IMU, force sensors. The simulated robot "sees" what a real robot would see.
• Domain randomization: automatically vary lighting, textures, object positions, and physics parameters so the robot generalizes to unseen environments.
• Synthetic data generation: generate millions of labeled training images without manual annotation. The GR00T-Dreams blueprint generated 780,000 synthetic trajectories (9 months of human demonstration data) in just 11 hours.

Isaac Lab (training)

GPU-accelerated reinforcement learning at massive scale. Key capabilities:

• Parallel training: run thousands of robot instances simultaneously on GPU clusters. What takes weeks in the real world takes minutes in simulation.
• Multi-physics: rigid bodies, soft bodies, fluids, cloth. Complex manipulation tasks like folding laundry or handling deformable objects.
• Isaac Lab 3.0: released in early access at GTC 2026. Runs on DGX-class infrastructure. Enhanced support for dexterous manipulation.
• Isaac Lab-Arena: simplifies environment composition and accelerates complex task creation. Evaluates tasks in parallel.

Isaac ROS (robot runtime)

The software that runs ON the robot during operation. CUDA-accelerated ROS2.

• Built on ROS2: the industry standard robotics framework. Millions of ROS developers can use NVIDIA-accelerated libraries directly.
• GPU-accelerated perception: object detection, segmentation, depth estimation, all running on the Jetson GPU. 10-100x faster than CPU-only ROS2.
• Navigation stack: autonomous navigation using camera and LIDAR data. Path planning, obstacle avoidance, SLAM.
• Manipulation libraries: grasp planning, motion planning for robotic arms.

Newton (physics engine)

Open-source GPU physics engine. Built in collaboration with Google DeepMind and Disney Research.

• Purpose: make simulation physics match real-world physics closely enough that sim-trained skills transfer to real robots.
• Capabilities: accurate collision detection, realistic object contact, stable simulation of complex systems with both rigid and flexible parts.
• GA: Newton 1.0 reached general availability in April 2026.
• Disney connection: Disney Research uses Newton for its robotic character platform (the Olaf BDX droids use NVIDIA Jetson + Omniverse + Newton).

5. Foundation models: GR00T and Cosmos

Isaac GR00T

The most ambitious piece of the NVIDIA stack. Foundation models for humanoid robots.

The concept: just as GPT is a general-purpose language model that can be fine-tuned for specific text tasks, GR00T is a general-purpose robot model that can be fine-tuned for specific physical tasks. It takes visual input (what the robot sees) and language input (what it is told to do) and outputs motor commands.

Evolution timeline:

• GR00T N1 (GTC 2025): first open foundation model for humanoid reasoning and skills. Cross-embodiment (works across different robot bodies).
• GR00T N1.5: generated using synthetic data from GR00T-Dreams blueprint in 36 hours. 40% performance improvement over N1 when combining synthetic and real data.
• GR00T N1.6 (CES 2026, January 2026): open reasoning VLA model. Unlocks full body control. Uses Cosmos Reason for better contextual understanding.
• GR00T N1.7 (May 2026): supports whole-body humanoid control via the Unitree G1 (UNITREE_G1_SONIC embodiment + GEAR-SONIC controller). The VLA predicts latent action tokens that a whole-body controller decodes into full-body joint commands including legs, arms, and hands. Single policy produces language-conditioned, coordinated manipulation and locomotion end-to-end.
• GR00T N (GTC 2026): extended to surgical robotics applications.

Key concept: VLA (Vision-Language-Action). GR00T N1.6+ is a VLA model. It sees, understands instructions, and acts. This is the architecture that makes "pick up the red cup" work without explicit programming for cups.

Code: GitHub | Docs: NVIDIA Developer

Cosmos world models

Cosmos is NVIDIA's family of "world foundation models." Think of them as the robot's imagination.

What they do: Cosmos models predict what will happen next in the physical world. If a robot pushes a cup toward the edge of a table, the world model predicts the cup will fall. This predictive ability is essential for planning actions and evaluating whether a planned action will succeed before executing it.

Key models (as of GTC 2026):

• Cosmos Transfer 2.5: converts real video to synthetic training data. Enables massive data augmentation from small real-world datasets.
• Cosmos Predict 2.5: predicts future states of physical environments. Enables robot policy evaluation in simulation.
• Cosmos Reason 2: open reasoning VLM. Enables machines to see, understand, and act in the physical world. The perception layer for GR00T N1.6+.
• Cosmos 3.0 (GTC 2026): first world foundation model unifying synthetic world generation, vision reasoning, and action simulation.

All Cosmos models are open and available on Hugging Face.

6. The competing thesis: LeCun, AMI Labs, and JEPA

Not everyone agrees that the NVIDIA approach (massive simulation pipelines, synthetic data, scaled compute) is the right path. The most credible dissent comes from Yann LeCun, who helped create the deep learning revolution and is now betting $1 billion that the entire industry has taken a wrong turn.

What happened

LeCun left Meta in late 2025 after 12 years leading FAIR (Fundamental AI Research). He founded AMI Labs (Advanced Machine Intelligence) in Paris with co-founder Alexandre LeBrun. In March 2026, AMI raised $1.03 billion in seed funding at a $3.5 billion pre-money valuation, the largest seed round in European history, backed by Bezos Expeditions, NVIDIA, Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and others including Tim Berners-Lee, Jim Breyer, Mark Cuban, Xavier Niel, and Eric Schmidt.

His departure was driven by a philosophical split. Meta pivoted resources to catch up in the LLM race (Llama 4). LeCun believed true machine intelligence requires understanding physical reality, not predicting the next token.

The core argument

LeCun's position: LLMs are sophisticated pattern matchers that produce plausible text without ever understanding the world. They predict the next word. Image generators predict pixels. Both operate in "surface space" and fail when applied to physical-world reasoning because the space of possible outputs is astronomically large (every possible future frame of video, every possible physical outcome).

His alternative: world models that learn abstract representations of reality and make predictions in "latent space." Instead of predicting every pixel of what happens next, the model predicts the abstract state of the world: "the cup will fall off the table" rather than generating a video of the cup falling. This is how LeCun believes babies learn physics: by observing patterns, not by generating predictions of every sensory detail.

Key quote: "I do not understand how you can even think of building an agentic system without that system having the ability of predicting the consequences of its actions."

JEPA (Joint Embedding Predictive Architecture)

JEPA is the technical architecture behind LeCun's thesis. It trains AI systems to predict abstract representations of future states in a learned embedding space, rather than predicting raw data.

The progression:

• I-JEPA: image-based JEPA. Learns to predict abstract image representations.
• V-JEPA 2 (June 2025): video-based JEPA. 1.2B parameters. Trained on 1M+ hours of unlabeled internet video. State-of-the-art on action anticipation (predicting what will happen 1 second in the future from video).
• V-JEPA 2-AC (Action-Conditioned): after pre-training on video, a second phase adds just 62 hours of robot control data. The model can then predict physical outcomes of robot actions, enabling planning and control.
• V-JEPA 2.1 (March 16, 2026): improved training recipe for high-quality, temporally consistent dense features. All tokens (visible and masked) now contribute to the self-supervised training loss. State-of-the-art robot navigation. 10x faster planning than previous approaches. Closes the gap with DINOv2 on dense, spatially precise features.

Real results, not just theory

V-JEPA 2 has been tested on physical robots at Meta's labs:

• Zero-shot robot control: 65-80% success rate on pick-and-place tasks with objects and environments the model had never seen. No retraining for new scenes.
• 30x faster than NVIDIA's Cosmos on major video benchmarks (Something-Something v2, Epic-Kitchens-100).
• 50% fewer parameters than comparable models with equivalent benchmark performance.
• State-of-the-art action anticipation: predicts what a person will do 1 second in the future from egocentric video, outperforming task-specific models.

The 62-hour training number is striking. GR00T required 780,000 synthetic trajectories (equivalent to 9 months of human demonstration data). V-JEPA 2 achieves competitive robot control from 62 hours of real robot data on top of video pre-training. This is the efficiency argument in action.

NVIDIA GR00T vs LeCun JEPA: head to head

The honest assessment

LeCun's credibility is beyond question: he invented convolutional neural networks, won the Turing Award, and built the research lab that produced many of the tools the industry uses today. The V-JEPA results are real and published. The $1B funding gives AMI Labs 4-5 years of runway.

But AMI Labs is months old and has shipped nothing commercially. NVIDIA has 110+ partners with deployed systems. The gap between "impressive research" and "robots doing useful work at scale" is wide. If JEPA's efficiency claims hold at scale, it could reduce the need for massive simulation infrastructure. If they do not hold, the approach remains a research contribution rather than an industry shift.

Both outcomes matter for Physical AI work. Either NVIDIA's full-stack approach wins and the skills around it remain central, or JEPA-style world models win and the skills shift toward efficient representation learning and minimal-data training. Perception, edge deployment, and governance skills are relevant either way.

Where governance sits in both

The System M paper (Dupoux, LeCun, Malik; arXiv 2603.15381, March 2026) proposed an A-B-M architecture for autonomous cognition. The "M" (Meta-learning system) orchestrates how the robot learns and adapts. What is missing from both NVIDIA's and LeCun's approaches: governance. Who decides what the robot should learn? What boundaries constrain its autonomous behavior? How do you audit a world model's internal representations? These are the same questions software AI governance is asking, with physical consequences.

Links

• AMI Labs site
• V-JEPA 2 (Meta AI)
• V-JEPA 2 paper (arXiv)
• System M paper (arXiv)
• MIT Tech Review interview

7. The open-source ecosystem: Hugging Face and beyond

Why Hugging Face keeps showing up

Hugging Face is to AI what GitHub is to software. With over 1 million repositories and 5 million+ researchers and developers, it is the infrastructure layer the AI ecosystem runs on. When a new model is released (by NVIDIA, Google, Meta, Alibaba, or anyone else), it shows up on Hugging Face first.

Their robotics push is recent and aggressive. In early 2025, Hugging Face acquired Pollen Robotics (makers of Reachy Mini), creating a vertically integrated open-source robotics stack: software (LeRobot) + hardware (Reachy Mini) + model hosting (Hugging Face Hub) + community. Robotics datasets on the platform grew from 1,145 in 2024 to 26,991 in 2025, making robotics the single largest dataset category on the entire platform.

The Reachy Mini app store (May 6, 2026)

Hugging Face launched an open-source app store for Reachy Mini on May 6, 2026, opening with 200+ apps from 150 different creators. Around 10,000 Reachy Mini devices were already shipped or in transit at launch, with 3,000 shipped in the week prior. The hardware ranges from $299 (Lite, tethered) to $449 (Wireless, RPi 5 with battery).

Apps require no coding knowledge to build or use. Examples at launch include an office receptionist app built in under two hours, baby-monitor-style apps, cooking assistants, and a distraction tracker. Hugging Face's "ML Intern" agent can generate and refine apps. Every app is an open-source repository that anyone can fork, audit, or extend. This mirrors the iOS App Store model but for robot intelligence.

The NVIDIA partnership makes this ecosystem more powerful. As of CES 2026, NVIDIA Isaac models and libraries integrate into LeRobot. A hobby arm running LeRobot connects to the same model ecosystem that Figure's humanoid fleet uses. The workflow: collect demonstration data with LeRobot, train with Isaac Lab, simulate in Isaac Sim, deploy on Jetson.

Hugging Face LeRobot

• Framework: PyTorch-based. Covers data collection, training, simulation, and deployment.
• Hardware compatibility: SO-ARM100/101 robotic arms, Reachy Mini, any LeRobot-compatible robot.
• NVIDIA integration: Isaac open models and libraries integrated into LeRobot as of CES 2026.
• Approach: imitation learning (demonstration-based) and reinforcement learning.

Site: Hugging Face LeRobot | Code: GitHub

Other platforms and tools

Key open hardware projects

8. Autonomous vehicles: Tesla, Waymo, and the road

This is the most visible and commercially advanced application of Physical AI. Self-driving vehicles represent billions of dollars in real-world deployment, millions of miles of real-world data, and some of the most complex AI challenges in existence.

Tesla: the Physical AI company

Tesla no longer describes itself as a car company. In its 2025 annual report, Tesla described its transition "from a hardware-centric business to a physical AI company." Three Physical AI products are now active or launching:

Full Self-Driving (FSD): Tesla's autonomous driving software, now at version 14.3 (April 2026). Uses end-to-end neural networks trained on data from millions of vehicles. The v14.3 update upgraded the reinforcement learning stage for edge cases, enhanced the vision encoder for low-visibility scenarios, and rewrote the AI compiler to cut inference latency by up to 20%. On the Q1 2026 earnings call, Musk said unsupervised FSD would reach customer vehicles "probably Q4" of this year, though Tesla's FSD timelines have been consistently optimistic.

Cybercab (robotaxi): a purpose-built autonomous vehicle with no steering wheel or pedals. Continuous production began at Giga Texas in April 2026 (the first unit rolled off in February). Tesla designed the Cybercab to comply with all existing FMVSS standards on its own, so it is not subject to NHTSA's 2,500-vehicle annual production cap for autonomous vehicles. Robotaxi service expanding to seven cities in H1 2026 (Dallas, Houston, Phoenix, Miami, Orlando, Tampa, Las Vegas) on top of existing service in Austin. Showcased at the F1 Miami Grand Prix Fan Fest (April 29 to May 3, 2026) in a glass display case.

Optimus (humanoid robot): Gen 3, the first version designed for mass production, was unveiled in Q1 2026 with an updated hand design. The AI5 inference chip was taped out on April 15, 2026, and will deploy to Optimus robots and supercomputer clusters first, not Tesla vehicles. Optimus integrates xAI's Grok for conversational AI while using FSD-derived neural networks for physical movement. Musk has stated plans to produce 1 million Optimus units annually, and SpaceX aims to send Optimus robots to Mars on Starship flights in late 2026.

Why Tesla matters here: they have the largest real-world driving dataset on Earth. The FSD neural networks are the most battle-tested Physical AI systems in consumer deployment. The Optimus-FSD convergence (same AI stack powering both cars and humanoid robots) represents a unique approach where vehicle autonomy and robot autonomy share a single development pipeline.

Tesla AI: tesla.com/AI

Waymo (Alphabet)

The market leader in fully autonomous ride-hailing. Waymo operates commercial robotaxi services in multiple US cities. Their vehicles use a sensor suite (LIDAR, cameras, radar) and proprietary AI for fully driverless operation (no safety driver). Waymo holds the most advanced autonomous vehicle permits in the US.

waymo.com

Other autonomous vehicle companies

The data advantage

What sets autonomous vehicles apart from other Physical AI domains: they generate vast quantities of real-world training data continuously. Every Tesla on the road feeds data back. Every Waymo ride adds to the dataset. This creates a compounding advantage: more data produces better models, better models attract more users, more users generate more data. This flywheel does not exist in humanoid robotics yet, which is why vehicle autonomy is years ahead of robot autonomy in real-world deployment.

9. Companies building Physical AI

The NVIDIA partner ecosystem

As of GTC 2026 (March 2026), NVIDIA has 110+ robotics partners. The major categories:

Robot brain developers (generalized AI for any robot body)

• Skild AI: generalized robot brains using Cosmos world models and Isaac simulation.
• FieldAI: same approach as Skild, using Cosmos for data generation.
• World Labs: using Isaac Sim to validate generative world models.
• Generalist AI: using Cosmos to explore synthetic data generation.

Humanoid robot companies

Industrial robot giants

Surgical robotics

Other notable deployments

Source: NVIDIA GTC 2026 robotics roundup.

10. Medical and surgical AI

Surgical robotics is one of the most mature and commercially proven applications of Physical AI. The global surgical robotics market is projected to reach $14 billion by 2026.

Surgical robotics leaders

Beyond the operating room

Rehabilitation and exoskeletons: AI-driven robotic exoskeletons (Ekso Bionics, ReWalk) enable paraplegic patients to walk again. ML-controlled actuators adjust resistance based on patient progress, creating personalized rehabilitation plans. This area bridges Physical AI with wearable technology.

Smart prosthetics: machine-learning-controlled prosthetic limbs learn the wearer's movement patterns and adapt. Advances in neural interfaces are connecting prosthetics directly to the nervous system for near-natural control.

AI diagnostics: medical imaging AI (radiology, pathology, dermatology) using computer vision to detect disease. Closer to software AI than Physical AI, but the diagnostic models increasingly run on edge hardware in clinics and hospitals.

The regulatory landscape for medical AI

The FDA has been actively developing frameworks for AI in medical devices:

• January 2025: FDA published draft guidance on AI-enabled Device Software Functions (DSF), recommending model descriptions, data lineage, bias analysis, and human-AI workflow documentation.
• 2026 agenda: FDA finalizing lifecycle management rules for AI-enabled devices.
• EMA + FDA (January 2026): jointly published 10 guiding principles for good AI practice in medicines development.
• Key classification: real-time surgical AI predictions fall into Software as a Medical Device (SaMD). The line blurs when AI influences a robot's physical actions.

11. The defense and autonomous sector

This sector represents a massive and rapidly growing application of Physical AI. The skills overlap with commercial robotics is near-total: perception, autonomous navigation, edge AI, reinforcement learning, sim-to-real transfer.

Axon (Nasdaq: AXON)

What they are: global leader in public safety technology. Evolved from body cameras and TASERs into a full AI-powered public safety platform.

Scale: Q1 2026 revenue $807M (up 34% YoY, 9th consecutive quarter of 30%+ growth). AI revenue surged 700%. AI bookings rose 140% YoY. Annual recurring revenue $1.5B (up 35%). Platform Solutions revenue up 95% YoY to $111M. Management raised full-year 2026 guidance to 30-32% growth. Market cap ~$30B.

Product ecosystem:

• TASER energy devices (TASER 10 adopting at 2x the rate of TASER 7).
• Body-worn, fixed, and in-car cameras (Axon Body 4).
• Axon Vision: AI-powered real-time awareness from live video (CCTV analysis at scale).
• Axon Vehicle Intelligence: AI license plate recognition with vehicle descriptors.
• Dedrone: counter-drone detection and response.
• DFR (Drone First Responder): autonomous drones dispatched ahead of ground units.
• Draft One: AI-automated police report writing from body camera footage.
• Axon Evidence: cloud-based digital evidence management.

Site: axon.com | Q1 2026 earnings

Overland AI (private, Seattle)

What they are: autonomous ground vehicles for the U.S. military. Spun out of the University of Washington in 2022.

Scale: raised $100M in February 2026 (led by 8VC). 100+ employees. Total funding $140M+.

Product stack:

• OverDrive: autonomy software stack. Off-road navigation in GPS-denied, comms-denied environments. Computer vision, LIDAR, proprietary AI models.
• SPARK: modular sensor and compute hardware kit. Mounts on any vehicle to give it autonomous capability.
• OverWatch: command and control platform. One operator controls multiple autonomous ground vehicles.
• ULTRA: purpose-built autonomous ground vehicle platform.

Operational deployments (real, not demos):

• 82nd Airborne Division, JRTC (April 2026): four ULTRA vehicles integrated into a month-long combat training rotation. Reduced distribution timelines by ~50%.
• USMC ROGUE Fires (April 2026): integrated OverDrive onto the Marine Corps Remotely Operated Ground Unit for Expeditionary Fires.
• Selected for U.S. Army UxS Autonomous Maneuver Program to evaluate autonomy on Infantry Squad Vehicles.

Anduril Industries (private, $61B valuation)

What they are: the largest defense tech startup in the world. Founded 2017 by Palmer Luckey. Building autonomous weapons systems, drones, naval vessels, and the software platform that connects them all.

Scale: $5B Series H on May 13, 2026 led by Thrive Capital and Andreessen Horowitz, doubling valuation to $61B. Expected 2026 revenue ~$4.3B (doubling from 2025). The U.S. Army awarded Anduril a firm-fixed-price enterprise contract with a $20B ceiling for Lattice integration (10-year enterprise agreement). Also signed a framework agreement with the U.S. Department of War to mass-produce at least 3,000 Barracuda-500M cruise missiles over three years. Total funding $11B+.

Product ecosystem:

• Lattice: autonomous software platform and network. The "operating system" for the battlefield. Connects and powers thousands of systems, analyzes sensor data, acts autonomously under human supervision.
• Fury (YFQ-44A): unmanned autonomous fighter jet for the U.S. Air Force Collaborative Combat Aircraft (CCA) program.
• Roadrunner: jet-powered VTOL autonomous air vehicle. Reusable. Intercepts and destroys drones and aerial threats.
• Ghost: autonomous small UAS for ISR.
• Altius: tube-launched loitering munitions. Deployed in Ukraine.
• Barracuda-500M: surface-launched cruise missiles (3,000+ over three years).
• Autonomous Surface Vessels: being built at the old Foss Shipyard in Seattle for the U.S. Navy MASC program.
• Golden Dome: space-based missile defense system (contract announced May 2026).
• Arsenal-1: hyperscale manufacturing campus in Ohio. $1B investment. 5M+ sq ft. 4,000+ workers.

Shield AI (private, $12.7B valuation)

What they are: autonomous piloting software for military aircraft. "The company that builds the pilot, not the plane."

Scale: $1.5B Series G (March 2026) at $12.7B valuation. Projected 2026 revenue $540M+. Total funding $3.6B+.

Core technology:

• Hivemind: autonomous piloting software. Flies aircraft without human input or GPS. Selected for the U.S. Air Force CCA prototype program (February 2026).
• V-BAT: VTOL drone. Replicator-eligible (DoD mass production program).
• X-BAT: stealth autonomous combat aircraft. Flight testing scheduled 2026.

The Anduril-Shield AI dynamic: for the CCA program, Anduril builds the airframe (Fury) and Shield AI provides the autonomy software (Hivemind). The Air Force deliberately split hardware and software to avoid single-vendor lock-in. Anduril supplies the body, Shield AI supplies the brain.

Other companies

The funding context

• Pentagon FY2026 budget: ~$8.9B for autonomous and collaborative platforms.
• Replicator program: ~$1B across two tranches for mass-produced autonomous systems.
• CCA program: ~$1B in Air Force RDT&E (the single largest line item).
• Anduril: $11B+ total funding, $61B valuation, $20B Lattice contract.
• Shield AI: $3.6B+ total funding, $12.7B valuation.
• Overland AI: $140M+ total funding.

The Pacific Northwest connection

Seattle and the surrounding area is emerging as a major hub for defense tech and Physical AI: Overland AI (autonomous military ground vehicles, UW spinoff), Anduril (autonomous naval vessels at the Seattle shipyard), Echodyne (radar), Armada (Bellevue, portable AI data centers), Amazon Robotics (warehouse/logistics), and NVIDIA ecosystem partners.

12. Space exploration

Space robotics is one of the oldest and most demanding applications of autonomous AI. Communication delays (up to 24 minutes to Mars one way) mean space robots must operate autonomously, without real-time human control.

NASA's AI robotics fleet (2026)

Perseverance Mars Rover: in December 2025, NASA's Perseverance completed the first AI-planned drives on another world. Generative AI analyzed high-resolution orbital images of Jezero Crater and generated waypoints that kept the rover away from hazardous terrain. Previously, human route planners handled this complex decision-making task.

Astrobee (ISS): three cube-shaped free-flying robots (Honey, Queen, Bumble) operate inside the International Space Station. They inventory supplies, record video, monitor systems. In late 2025, Stanford researchers demonstrated the first AI-controlled robotic navigation on the ISS, achieving 50-60% faster autonomous movement planning.

CIMON: an AI-powered robot about the size of a bowling ball, tested on the ISS for its ability to control other free-flying robots. Robot-to-robot communication and autonomous task delegation in space.

Robonaut 2: a humanoid robot at the ISS with 350+ sensors, 38 processors, and hands with 12 degrees of freedom. Can execute tasks autonomously. A precursor to the humanoid robots that will eventually work alongside astronauts on the Moon and Mars.

Fly Foundational Robots (FFR): a NASA mission launching a robotic arm capable of dexterous manipulation, autonomous tool use, and walking across spacecraft structures in zero gravity. Developed by Motiv Space Systems. The stepping stone to robots that assemble solar arrays, refuel satellites, and build lunar habitats.

SpaceX + Tesla Optimus on Mars

SpaceX aims to launch a small uncrewed fleet of Starships toward Mars in late 2026, carrying Tesla Optimus humanoid robots to explore the Red Planet. Whether this timeline holds is debatable, but the convergence of SpaceX launch capability and Tesla humanoid robotics represents a path where Physical AI operates on another planet within the next few years.

Why space robotics matters

Space robotics operates under extreme constraints: limited compute, harsh environments, communication delays, no ability to physically intervene when something goes wrong. The AI must be resilient, autonomous, and power-efficient. These constraints mirror edge AI challenges on Earth (limited compute, real-time requirements, no cloud fallback). Skills in autonomous navigation, computer vision on constrained hardware, and fault-tolerant AI transfer directly between space and terrestrial robotics.

13. Government regulation and policy

Physical AI regulation is fragmented, evolving rapidly, and presents a significant governance gap. This is one of the most under-developed and under-discussed areas in AI policy today.

EU AI Act and the May 2026 Omnibus simplification

The EU AI Act is the world's first unified law governing AI systems. It entered into force on August 1, 2024, with full applicability originally scheduled for August 2, 2026.

On May 7, 2026, the EU Council and Parliament agreed on the "Digital Omnibus on AI" (part of the Omnibus VII simplification package), making significant changes:

• High-risk AI deadlines deferred: obligations applicable to high-risk AI systems subject to existing EU sectoral legislation (Article 6(1) coverage) now take effect on August 2, 2028 rather than 2026.
• Machinery overlap resolved: machinery products now only need to comply with sectoral safety rules instead of both the AI Act and sectoral rules. Direct impact on industrial robotics, factory automation, and surgical equipment.
• Transparency timeline: grace period for providers to implement transparency solutions for AI-generated content reduced from 6 months to 3 months, with the new deadline December 2, 2026.
• Formal adoption expected before August 2, 2026.

Key provisions still affecting Physical AI:

• Autonomous vehicles: AI systems in autonomous vehicles qualify as high-risk under Annex I (product safety component). Subject to conformity assessments, technical documentation, and human oversight requirements.
• Surgical robots: real-time surgical AI falls into Software as a Medical Device (SaMD). High-risk classification. Rules for AI embedded in regulated medical products apply from August 2028.
• Military exemption: the EU AI Act exempts AI systems designed for military, national defense, and national security purposes. This creates a regulatory gap where defense AI (autonomous weapons, military drones) operates outside the Act's framework entirely.
• Enforcement reality: non-compliance can cost up to 7% of global annual revenue.

EU AI Act framework | May 7, 2026 Council press release

US approach (fragmented, sector-specific)

The US has no comprehensive federal AI regulation equivalent to the EU AI Act. Instead, regulation happens through existing agencies:

State-level activity is filling federal gaps:

• Colorado AI Act (effective February 2026): most comprehensive US state AI legislation. Regulates high-risk AI in employment, housing, education, financial services.
• California: multiple bills addressing automated employment decisions, insurance algorithms, political deepfakes.
• NYC Bias Audit Law: mandates annual bias audits for automated employment decision tools.

Autonomous weapons regulation (global)

The most contested governance area in Physical AI. International efforts to regulate autonomous weapons have been ongoing for nearly a decade with limited progress:

• UN Convention on Certain Conventional Weapons (CCW): the Group of Governmental Experts on Lethal Autonomous Weapons Systems (LAWS) has been meeting since 2014. Mandate extended to 2026. A "rolling text" summarizing possible regulatory elements exists, but geopolitical tensions have blocked binding agreements.
• UN Secretary-General: has called for a legally binding treaty to prohibit LAWS that function without human control, to be concluded by 2026.
• International Committee of the Red Cross (ICRC): urges states to negotiate new legally binding rules on autonomy in weapons systems, with both prohibitions and regulations.
• Stop Killer Robots: global coalition campaigning against autonomous weapons. Active in 100+ countries.
• The EU gap: the AI Act's military exemption means defense AI developed in Europe operates outside civilian AI governance.
• The US reality: companies like Anduril, Shield AI, and Lockheed Martin are deploying autonomous weapons systems with human-in-the-loop oversight, but the degree of autonomy is increasing. The $20B Lattice contract (Anduril + US Army) puts autonomous decision-support across the military.

Industry response: NemoClaw (NVIDIA's first governance product)

NVIDIA announced NemoClaw at GTC 2026 as a security and governance layer for AI agents. It monitors the reasoning process of AI agents in real time and ensures their actions align with pre-defined safety guardrails. The key phrase from NVIDIA: it "inspects the intent of the AI's logic." This is the first product-level acknowledgment from a major Physical AI platform that governance of autonomous systems requires inspecting intent, not just monitoring actions.

NemoClaw works with AgenticROS, an open-source project that connects AI agents (Claude Code, Claude Desktop, Gemini, MCP) to ROS2 robots via natural language. The combination means you can now talk to a physical robot through Claude, and NemoClaw monitors whether the robot's AI is staying within its safety boundaries. This is exactly the governance-at-the-intent-layer problem.

Both are early. NemoClaw is a security product, not a full governance framework. It monitors and sandboxes; it does not address accountability, liability, auditability, or the broader policy questions. But it validates the direction.

AgenticROS GitHub | agenticros.com

14. Skills that matter across Physical AI

The Physical AI stack rewards engineers who can move across layers: perception on the camera, planning in the model, control on the actuator, governance over the whole thing. The skills below show up in roles at every company in this document.

Role types

15. Key resources and links

Primary (start here)

• Jetson AI Lab: hands-on tutorials for GenAI on Jetson. THE starting point.
• Jetson Orin Nano Super (buy): the hardware entry point.
• NVIDIA Robotics Platform: overview of the full platform.
• Isaac GR00T GitHub: open-source humanoid foundation model.
• LeRobot: open-source robotics framework.
• SO-ARM100 GitHub: open-source robotic arm.

NVIDIA software documentation

Isaac Overview · Isaac Sim · Isaac Lab · Isaac ROS · Isaac GR00T · Newton Physics · Cosmos Models · Metropolis · DeepStream SDK · JetPack SDK

Developer resources

Jetson Developer Portal · Jetson Forums · Jetson AI Lab Discord · NVIDIA Robotics Blog · Jetson Store

Companies (quick reference)

Axon · Overland AI · Anduril · Shield AI · Figure · Agility Robotics · Boston Dynamics · Unitree · Skydio · 1X Technologies · AMI Labs

Key recent articles (verified May 2026)

16. Full glossary

AI and computing

Robotics

NVIDIA-specific

LeCun / AMI Labs

Autonomous vehicles

Defense