EMNLP 2026: Auto-Robotist Accumulates Reusable Design Knowledge for AI Robots

Researchers from the University of Michigan Ann Arbor have developed Auto-Robotist, a system that integrates structural search with controller training to accumulate reusable design knowledge. The framework uses large language models to generate 'design notes' during the search process, guiding subsequent designs by recording patterns of success and failure. Experiments in the EvoGym soft-body robotics simulation platform show that Auto-Robotist outperforms standard genetic algorithms across seven tasks, achieving higher final scores and converging faster. The system successfully transfers design rules from a 5x5 grid to a larger 10x10 design space without simply scaling up specific shapes, demonstrating that the accumulated knowledge helps robots maintain stable structures like walking frames and load-bearing trunks.

Researchers at the University of Michigan propose Auto-Robotist, an agent that combines structural search with controller training to design soft robots. Using a large language model, the system generates and maintains a continuously updated skills library from simulation results in EvoGym. Experiments show it outperforms standard genetic algorithms on seven tasks across 5x5 and 10x10 grid spaces, achieving higher scores and faster convergence while learning from both successes and failures.