Physical AI FAQ

A robot is a machine: motors, sensors, a mechanical structure. It can be entirely devoid of intelligence: most industrial robots installed worldwide replay pre-programmed trajectories without perceiving anything around them. Physical AI is the intelligence layer that lets a machine perceive, decide and adapt. A robot becomes a physical AI system when it copes with the unexpected: grasping a misplaced part, avoiding an operator, adjusting its motion on the fly. The robot is the body; physical AI is the brain. And that brain can power more than robots: vehicles, drones, machine tools.

The three terms overlap without being interchangeable. Robotics is the engineering discipline that builds the machines: mechanics, actuators, control. Embodied AI is a research current holding that genuine intelligence must be learned through a body interacting with the world. Physical AI is the industry umbrella term, popularized notably by NVIDIA, covering every AI system that operates in the real world: robots, but also autonomous vehicles, drones and intelligent industrial equipment. In short: robotics builds the body, embodied AI theorizes learning through embodiment, and physical AI names the market that results from both.

A VLA model is a neural network that takes in images (vision) and a natural-language instruction (language), and directly outputs robot commands (action). A typical instruction: put the blue parts in the left bin. Descendants of large multimodal models, VLAs are trained on robot demonstrations plus web-scale data, which gives them an unprecedented ability to generalize: performing task variants never seen during training. Well-known examples include Google DeepMind's RT-2, Physical Intelligence's π0, NVIDIA's GR00T and Figure's Helix. This is the building block that brings robotics closest to a ChatGPT-style way of working.

An LLM reads and produces text; a VLA sees, understands an instruction and produces motor actions. A VLA reuses the architecture and pre-training of LLMs (it often embeds one as its brain), but it runs in a real-time control loop, from 10 to 200+ Hz, where an LLM can afford seconds of latency. Its training data (robot demonstrations) is scarce and expensive, and its errors carry a physical, sometimes irreversible cost: safety is a design requirement, not an option. Our guide covers the full comparison.

Ready for pilots, yes; for full-scale production, not yet. Several carmakers and logistics operators are testing humanoids on tightly scoped tasks: moving totes, tending machines, simple sorting. These pilots work, but at slower cycle times than a human operator, under human supervision and sometimes with standby teleoperation. The remaining hurdles are reliability over thousands of hours, safety when working alongside people, and the full cost of integration. 2026 marks the transition from demos to the first limited deployments: the right time to run experiments, not yet the time to bet everything on them. Track this shift in our humanoid production index.

The range is very wide. The entry level is the Unitree G1, starting at roughly $16,000 in its base configuration, aimed at research and education; versions with dexterous hands and the full SDK exceed $70,000. Western industrial-grade platforms trade above $150,000, often leased or offered as robot-as-a-service. Tesla is targeting $25,000 to $30,000 for Optimus eventually, a stated goal that remains unproven. Keep in mind that hardware is only part of the bill: integration, safety engineering and maintenance can cost more than the robot itself. Compare these models in our robots comparator.

Sim2real (simulation to reality) means training a physical AI system in a virtual world before transferring it to a real robot. Simulation provides millions of fast trials with no breakage or danger, and endless environment variations. The challenge is the gap between simulated and real physics: friction, lighting, imperfect sensors. Teams close it with domain randomization (deliberately varying parameters to force robustness) and with increasingly faithful simulators such as NVIDIA Isaac Sim. Mistral AI's Robostral Navigate, trained entirely in simulation, is a good illustration of how mature the approach has become.

Three clusters stand out. Infrastructure providers: NVIDIA dominates compute, simulation (Isaac, Omniverse) and open models (GR00T); Google DeepMind is pushing Gemini Robotics. Humanoid makers: Figure, valued at $39 billion in late 2025, Tesla, Boston Dynamics and Agility Robotics on the American side; Unitree and UBTECH on the Chinese side, with a price and supply chain advantage. Finally, robot foundation model labs such as Physical Intelligence and Skild AI. The United States concentrates the capital and the models, China the mass production. No single player has locked up the full stack yet.

Yes, though not at the center of the ring. Europe has neither America's mega-rounds nor China's industrial scale, but it fields serious contenders: Neura Robotics (Germany) on humanoids, PAL Robotics (Spain), a historic pioneer of the field, 1X (Norway) on home humanoids, Wandelbots (Germany) on no-code robot programming software, ANYbotics (Switzerland) on quadruped inspection, and Mistral AI (France), which released its first robotics model, Robostral Navigate, in July 2026. Above all, Europe owns the playing field: an exceptional density of factories and system integrators. Its winning card is industrial deployment, not the race for giant models.

Yes, but through a specific route, and the timeline has just shifted. A robot is first and foremost a machine: it falls under the EU Machinery Regulation 2023/1230, which governs CE marking and safety. Amendments adopted in 2026 (the digital omnibus package) postponed the high-risk obligations: December 2027 for Annex III use cases, and for AI embedded in machinery, requirements folded into the Machinery Regulation through delegated acts expected by August 2028. In practice: machinery compliance remains the immediate benchmark, while the AI layer firms up gradually. Follow the European Commission's official page.

Not with a humanoid. Start with a bounded, measurable use case with a short payback: automated inspection, internal transport with AMRs, machine tending, vision-based quality control. Then audit your prerequisites: connectivity, production data, in-house skills, safety requirements. Run a three-to-six-month pilot with success metrics defined up front (uptime, cycle time, error rate, total cost), and involve the operators from day one. Handle safety and CE marking at the design stage, not as an afterthought. Finally, favor interoperable platforms so you do not end up locked into a single vendor.