LocoFormer: Generalist Locomotion via Long-context Adaptation

Min Liu, Deepak Pathak, Ananye Agarwal · 2025

Read LocoFormer: Generalist Locomotion via Long-context Adaptation by Min Liu, Deepak Pathak, Ananye Agarwal (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-23745.txt.

Description

A controller tuned for one robot body usually fails on the next, so Min Liu and 2 co-authors trained a generalist instead. LocoFormer learns locomotion for legged and wheeled robots it has never seen, even without exact knowledge of their build. Two choices drive that flexibility. Training runs massive reinforcement learning on procedurally generated robots with heavy randomization.

The second choice is memory: context windows stretch across episode boundaries, orders of magnitude longer than usual. One deployed model handled varied robots through weight changes and motor failures. In extreme trials it even learned from falls in early episodes to control better in later ones.

Catalog record

CREATORMin Liu, Deepak Pathak, Ananye Agarwal
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-23745

Browse the full library · About the Tower of Babel