01
Sense
Cameras, force sensors, lidar, radar and other inputs capture the state of the machine and its environment.
Guide
What is Physical AI?
Physical AI is artificial intelligence embodied in machines that perceive their environment, make decisions and produce physical actions. The important unit is not a model or a robot alone, but the complete sensing, reasoning, compute and control loop.
Updated 30 September 2026
02
Reason
Models and policies interpret context, predict outcomes and select an action under physical and safety constraints.
03
Compute
Onboard processors, external computers or cloud infrastructure execute different parts of inference, control and training.
04
Act
Controllers, actuators and tools turn a policy into movement, manipulation or navigation, then feed the outcome back into the loop.
Physical AI is broader than robotics
Robotics describes the engineering of machines that sense and move. Physical AI focuses on the learned intelligence that lets those machines adapt, generalize and operate in less structured environments. The fields overlap, but they are not synonyms.
Embodied AI studies intelligence that emerges through interaction between an agent, a body and an environment. It is often used for research and foundation-model work.
Generative Physical AI applies generative models to world understanding, simulation, planning or action generation. It remains software; it does not identify the processor on which a robot runs.
Autonomy is a spectrum. A system may be teleoperated, supervised, autonomous for one task or able to generalize across tasks. The Atlas records product and deployment evidence rather than treating every demonstration as commercial autonomy.
Software and hardware are different layers
A robot can use several processors and several software systems. A partnership, an investment or support for a software framework does not prove that a particular chip is installed onboard.
AI and software stack: models, policies, runtimes, robot operating systems, simulation tools and cloud services.
Compute hardware: the documented silicon, module or computer used by a named product or configuration, together with its role and evidence level.
The Physical AI landscape
The Atlas groups companies by the physical system or enabling layer they primarily build. Each sector page reports the current company count, disclosed funding and product maturity represented in the dataset.
Humanoid Robotics
Companies developing general-purpose and task-specific humanoid robots, from prototypes and pilots to systems shipping into real workplaces.
Autonomous Vehicles
Companies building autonomous driving systems, robotaxis and self-driving vehicle platforms. This category is tracked separately from the core funding comparison.
Industrial Robotics / Cobots
Industrial robot and collaborative robot makers automating manufacturing, inspection, handling and other production workflows.
Drones / Aerial
Aerial robotics companies building autonomous drones and drone platforms for commercial, industrial, public-safety and defence applications.
Embodied AI / Foundation Models
Companies developing foundation models, robot learning systems and general-purpose autonomy software that connect perception, reasoning and physical action.
Edge AI Hardware
Semiconductor and edge-computing companies providing the processors, modules and systems that run physical AI workloads close to machines and sensors.
Logistics / AMR
Autonomous mobile robot and logistics automation companies serving warehouses, fulfilment centres, factories and material-handling operations.
Agricultural Robotics
Robotics companies applying autonomy to farming, including planting, weeding, spraying, harvesting, monitoring and autonomous machinery.
Maritime Robotics
Companies building autonomous surface and underwater systems for ocean data, inspection, security, research and offshore operations.
Explore the evidence
Use the guide as a starting point, then move into the underlying companies, market structure and sourced real-world examples.