Robot Foundation Models

RT-H separates language-motion choices from detailed robot actions. Learn when a language correction addresses the error and when a trajectory demonstration is needed.

A promising point on an object is only part of an action. Where2Act separates actionability, orientation proposals and scores for articulated interactions.

PerAct uses a voxel representation; Act3D refines spatial search. Compare their precision and coverage decisions without mixing benchmark conditions.

RoboPoint predicts a useful image location. Valid depth, camera calibration and a defined motion primitive are still needed to make it a robot target.

Code as Policies needs more than valid Python: check every generated helper, coordinate convention and feedback loop before connecting robot APIs.

A frozen Sparsh encoder can reduce label demand for a new touch task, but compare it with end-to-end training on the same sensor, preprocessing and held-out episodes.

Map Octo’s cameras, proprioception, history masks, task conditioning and action head to a new robot before tuning or rollout.

FoAR gates force/torque information around predicted contact. Diagnose the first contact transition before adding an always-on force stream to a robot policy.

A low keypoint-constraint cost can hide tracking, path, collision or execution failures. Diagnose the first boundary that breaks before changing the planner.

A fresh camera frame can still hide a target’s direction or a robot’s task phase. Use a bounded rollout check to separate missing temporal context from action delay.