AI BriefingLG AI Research presents LOAM at ICML 2026: LLMs design RL observation and action spaces for 3x faster learning
·2026.10.26 09:00
Key point
The method, presented at ICML 2026, generates observation and action models via LLMs to reconfigure the agent-environment interface, outperforming human-designed baselines in locomotion and manipulation tasks.
Details
- LG AI Research introduced LOAM (Large Language Model for Observation and Action Models) at ICML 2026, a framework where LLMs automatically design the observation and action spaces for reinforcement learning agents rather than relying on manual human design.
- In HumanoidBench locomotion tasks (stand, walk, run), LOAM reached high return more than three times faster than the human-designed FastTD3 baseline, demonstrating that LLM-generated interfaces significantly accelerate policy learning.
- LOAM-Race, a variant using Bayesian Ridge Regression, dynamically allocates training resources to the most promising design candidates, achieving higher final performance on complex manipulation tasks like Cube and Insert_small by efficiently exploring multiple LLM-generated options within a single training budget.
- The approach generalizes across different simulation environments and robot embodiments, including the Franka robot arm and Allegro Hand in Isaac Lab, where LOAM-Race achieved the highest final performance on the Repose-Cube task involving over 1,100-dimensional raw observations.
Methodology and Scope
- LOAM distinguishes between the raw environment space (all sensor data and control commands) and the agent space, using LLMs to generate Python functions that transform raw inputs into task-relevant observations and map policy outputs to executable actions.
- The system requires only task goals, robot structure, and variable meanings as input, allowing it to be adapted to new domains without modifying the underlying reinforcement learning algorithm.
- Future plans include expanding LOAM to LG's home and manufacturing robots, with a focus on automating the verification of physical validity and safety for generated models.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.