TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

robotics

Text Commands Reached the Humanoid Control Loop

ADAPT pairs a diffusion action prior with residual reinforcement learning for balance and smooth command changes.

Published Updated Story ID: mp-2026-09-02-016
Read the complete editionStory JSON

Summary

ADAPT pairs a diffusion action prior with residual reinforcement learning for balance and smooth command changes.

Instead of generating a kinematic clip for a separate tracker, the controller responds continuously to changing language while executing whole-body motion. A frozen diffusion prior supplies diverse skills and a lightweight residual policy improves robustness and prompt transitions. The abstract reports robust execution and reuse for downstream adaptation without presenting a universal safety claim.

Why it matters

ADAPT pairs a diffusion action prior with residual reinforcement learning for balance and smooth command changes.

Limits and context

No additional limitation was separately recorded.

Key claims

  1. ADAPT pairs a diffusion action prior with residual reinforcement learning for balance and smooth command changes.

    Evidence: source-2026-09-02-018

Sources

  1. arXiv preprint 2609.00677arXiv · primary research

Corrections

No corrections have been recorded for this story.