Robot-Foundation Imperialism

From the AI Research Compass wiki

Robot-Foundation Imperialism: One generalist policy should control every robot body, from arms to quadrupeds to humanoids, trained on pooled cross-embodiment data.

Beliefs

Robot-Foundation Imperialism wants to do for robots what pretraining did for text: pool data from every lab, every arm, every gripper and every mobile base into one corpus, and train one policy that runs on all of them. Each embodiment is a province with its own customs (joint counts, camera placements, control rates), but the empire insists they share a common language of manipulation and locomotion. Specialist per-robot policies are seen as feudal fragmentation.

It occupies the top-right corner of the chart: almost maximally World, because it lives entirely in physical action, and maximally Monolith, because one model must cover every body. Its paradigm level is Mainstream: it carries the Transformer-plus-scale recipe straight into robotics. Cross-embodiment training at scale is young, and the doctrine is betting that robotics will follow the same scaling path as language and vision.

Arguments

  • Pooling datasets across many robot types has shown positive transfer, with generalist policies outperforming policies trained on a single robot’s data.
  • A shared policy amortizes the enormous cost of robot data collection across the whole field.
  • Open generalist policies can be fine-tuned to new embodiments with modest data, lowering the barrier for new hardware.

Critiques

  • Embodiments differ in kinematics, sensing and dynamics so much that a single action space can be a lowest common denominator.
  • The pooled data is still tiny next to web text or video, so “foundation” may be premature.
  • Morphology-focused rivals argue the body itself does much of the computation, so one brain for every body misses the point.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ +1.00, y ≈ +0.96, paradigm ≈ −0.45.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data