FROM PROGRAMS
TO POLICIES.

A library of routines does not scale

A hand-authored routine can be precise, repeatable and safe. It can also be brittle. Move the object, change the lighting or ask for a slightly different goal and the routine may no longer apply.

Learned policies approach the problem differently. They infer an action from the observation, the goal and patterns acquired during training. The promise is adaptability; the engineering problem is keeping that adaptability grounded in a real body.

Vision, language and action

A vision-language-action model links what a robot sees, what a person asks and what the machine should do next. It may recognise an object, understand a spatial relationship and produce a sequence of motion targets.

The output is not intelligence floating above the hardware. Faster control layers still have to translate it into joint-level behaviour, maintain balance and enforce limits.

Google DeepMind positions Gemini Robotics as a family of models that add physical action and embodied reasoning to multimodal AI. NVIDIA describes GR00T as an open foundation-model family for generalist humanoid reasoning and skills. These are development programmes and platform claims—not proof that general-purpose autonomy is solved.

Robots need experience before they need scale

Physical data is expensive. A robot moves in real time, hardware wears, mistakes can damage the platform and a person often has to reset the scene. Developers therefore combine direct teleoperation, human-motion demonstrations, simulation, synthetic variation and autonomous trials.

Figure’s Project Go-Big describes a system intended to learn visual representations of human activity from internet video, while its Helix programme uses robot demonstrations and full-body policy learning. The strategic idea matters even where results remain company-reported: broad visual knowledge may become more useful when connected to scarce robot-action data.

The scarce resource is grounded experience: observation connected to action and consequence.

The same idea must survive a different body

A useful robotics model should not depend entirely on one arm length, one camera position or one factory cell. Cross-embodiment learning asks whether representations or skills can transfer while lower-level controllers adapt them to each machine.

This does not mean one universal policy will directly drive every motor. A wheeled manipulator and a full humanoid have different dynamics and safety envelopes. The more credible architecture is hierarchical: shared perception and task understanding above body-specific action and stabilisation.

Latency is part of intelligence

A capable model that responds too slowly can still produce a poor robot. On-device inference, compressed policies and fast feedback loops matter because the environment changes while an oversized system is thinking.

Watch reliability, not the next demo verb

The near future is likely to be less cinematic than “one model controls everything” and more useful: larger repertoires, faster adaptation and better recovery inside engineered boundaries.

• Long-horizon completion: can the robot finish without accumulating errors?
• Recovery: does it recognise failure and attempt a safe correction?
• Transfer: how much learning survives a change of platform?
• Local inference: what remains functional with strict latency?
• Verifiable safety: which constraints are enforced outside the learned model?

Generalist models do not remove control engineering, data curation or hardware design. They make the relationship between those disciplines more important.