Deep dive
The physical AI race is becoming a data race
Humanoid robotics is moving beyond hardware demonstrations toward a harder question: how can machines learn enough varied experience to work in the real world?
Hardware is only the visible half
Humanoid robots attract attention because their progress is easy to see. A new hand closes around an object, a machine walks through a factory, or a robot completes a household task. The visible movement can make the race appear primarily mechanical. In reality, the hardest long-term constraint may be the intelligence that lets one body perform many tasks across environments it has not encountered before.
A useful general-purpose robot must connect perception, language, memory, planning, balance, and manipulation in real time. It must understand that two visually different objects may be used in the same way, adapt a learned behavior to a new room, and recognize when uncertainty makes action unsafe. Hardware defines what movement is possible. Data and learning determine whether those possibilities become useful behavior.
The internet does not contain enough robot experience
Language models benefited from an enormous existing record of human expression. Robotics has no equivalent corpus of standardized physical experience. Demonstrations are expensive, robots differ in their sensors and bodies, and a failed action can damage an object or the machine itself. The industry has to create much of its training data while simultaneously building the systems that will learn from it.
That changes the economics of progress. A company needs fleets, teleoperation, simulation, evaluation environments, and a way to combine data gathered from different tasks. The resulting flywheel could become a major advantage: more deployed robots create more experience, which improves the model, which makes additional deployments possible. The risk is that data remains too narrow or too specific to one body to produce broad generalization.
Figure is integrating the full loop
Figure's Helix system is designed as a generalist vision-language-action model for humanoid control. The company describes a loop that connects perception, reasoning, and movement on the robot. Its newer work extends control across the full body and into longer sequences that combine walking, balance, and manipulation. This integrated strategy lets hardware and model development influence one another directly.
Figure has also emphasized data gathered from human-centered environments and the possibility of transferring learning from human video to robot behavior. That direction is significant because direct robot demonstrations alone may never scale quickly enough. If ordinary human activity can provide useful pretraining, the available data expands dramatically. The remaining question is how much physical precision can be learned from observation before robot-specific experience is still required.
Physical Intelligence is betting on transfer
Physical Intelligence is approaching the problem from the foundation-model layer. Its work aims to combine visual, language, web, demonstration, and robot experience data into models that can control different machines. The strategic idea is cross-embodiment learning: knowledge gained through one robot or task should make another easier to learn.
Successful transfer would change the structure of the robotics market. Hardware companies could begin with a broad base of learned capability rather than building every behavior internally. A shared model could improve as it absorbs experience from more embodiments. But abstraction has limits. A compact gripper, a dexterous hand, and a mobile humanoid do not experience the same forces or constraints. The model must generalize while preserving the precision required by each body.
Evaluation has to leave the laboratory
Robotics demonstrations compress time. Viewers see a completed task but not always the resets, intervention rate, environmental preparation, or performance after hours of operation. Commercial value depends on the unedited distribution: how often the robot succeeds, how it fails, how quickly it recovers, and whether the same behavior holds across lighting, clutter, objects, and people.
The field needs evaluation that reflects sustained deployment. Useful measures include intervention frequency, task completion over long horizons, damage and near misses, adaptation to variation, and the time required to teach a new skill. Benchmarks can help compare research, but customers will ultimately care about dependable work per dollar and the operational burden required to keep the fleet productive.
Manufacturing is part of the learning system
A data flywheel cannot develop without enough reliable machines in the world. That makes manufacturing quality, serviceability, power consumption, and component supply part of the AI strategy. A robot that spends too much time offline gathers less experience and creates less value. A design that cannot be produced consistently limits the variety of environments the model can encounter.
The strongest physical AI companies will treat deployment as research infrastructure. Each machine becomes both a product and a sensor for the next generation of capability. That creates tension around privacy, consent, data rights, and security, especially in homes and workplaces. The industry will need rules for what is collected, how it is used, and what control remains with the people sharing an environment with the robot.
The race is still open
No company has yet proven the complete economic case for a broadly capable humanoid at scale. The technical progress is real, but impressive movement is not the same as a sustainable product. Figure's integrated hardware-and-model strategy and Physical Intelligence's cross-platform model thesis represent two important approaches to the bottleneck.
The decisive advantage may belong to the organization that learns fastest from the widest useful experience while maintaining safety and hardware reliability. Physical AI will not be won by data volume alone. The data must contain the variation that teaches generalization, the feedback that identifies failure, and the structure that connects an observed action to a controllable machine. The next breakthrough may look less like a new robot and more like a better way for every robot to learn.
Primary sources
Figure Helix overview ↗Figure 03 ↗Figure Project Go-Big ↗Physical Intelligence π0 paper ↗