Auto-Robotist: LLM-Accumulated Design Knowledge Speeds Robot Co-Design Search 1.66x

新智元 · wechat · 2026-11-06

A University of Michigan team's Auto-Robotist, accepted to EMNLP 2026 main conference and open-sourced, combines robot morphology search with controller training by having an LLM distill trial-and-error experience into a reusable "skill library."

Motivation: On EvoGym, a 10×10 grid has 5^100 possible morphologies, and each candidate needs PPO controller training plus simulation — expensive, with GA baselines retaining no design knowledge.

Method: The skill library has three layers — L1 archetypes (useful structures), L2 rules (improvements and failure modes), L3 observations (supporting evidence). Each loop retrieves relevant experience to propose designs, validates via simulation (512k-step PPO training), then updates the library via Add/Diagnose/Merge. The LLM stays frozen; only the natural-language library evolves.

Results:

Cold-start 5×5 search beats GA on all seven tasks, with 1.66x average convergence speedup and leading 67% of evaluation steps
Cross-scale transfer to 10×10 beats GA on six of seven tasks (+1.45 average score gap)
Naively upscaling old designs fails (carrier score drops to 0.16); rules alone still win on some tasks
Ablations: removing GA mutation, L2/L3 rules, Diagnose, or Merge hurts the Pusher task by 24.0%, 23.8%, 21.6%, 14.0%

The final library held 5 archetypes, 12 positive and 17 negative rules — failures outnumbering successes. Validation remains limited to grid-based soft-robot simulation; real hardware is future work.

Original post →

Auto-Robotist: LLM-Accumulated Design Knowledge Speeds Robot Co-Design Search 1.66x

More from Embodied