Towards Adaptive Environment Generation for Training Embodied Agents

Teresa Yeo, Dulaj Weerakoon, Dulanga Weerakoon, Archan Misra · 2026

Read Towards Adaptive Environment Generation for Training Embodied Agents by Teresa Yeo, Dulaj Weerakoon, Dulanga Weerakoon, Archan Misra (2026) 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-2602-06366.txt.

Description

Robots and game agents trained in made-up worlds often fail in new ones, even when the worlds rhyme. Teresa Yeo and 3 co-authors argue that training worlds are usually generated blind, with no regard for what the agent currently finds hard. Easy worlds waste training time, and random variety is no cure.

Their proof-of-concept closes the loop. The system tracks detailed performance feedback, not just success or failure, and reshapes the next environments to press on the agent's weak spots. A controllable world representation lets difficulty be dialed along specific dimensions. The authors report that feedback-driven worlds produced more useful learning and better transfer to novel settings than open-loop generation.

Catalog record

CREATORTeresa Yeo, Dulaj Weerakoon, Dulanga Weerakoon, Archan Misra
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-06366

Browse the full library · About the Tower of Babel