Auto-Robotist Turns Robot Design Trials into Reusable Skills
The University of Michigan team behind Auto-Robotist says its open-source system uses large language models to propose robot-body changes, train controllers for candidates, and record simulation results as reusable design knowledge. Reported experiments on seven EvoGym soft-body robotics tasks found better final scores than the standard genetic-algorithm baseline, with the skill library also transferring from a 5×5 grid to a 10×10 design space. The evaluation was limited to simulated EvoGym tasks, so performance on three-dimensional designs and real hardware remains unverified.
Auto-Robotist offers a concrete approach to making expensive robot-design simulations reusable as evidence-backed rules for later searches, while its results remain limited to simulation.
新智元 2026-11-06 18:59 上海
新智元报道
AI能否像工程师一样,在反复设计机器人时积累经验?
机器人的身体结构与运动控制紧密相连:驱动器放在哪里、身体怎样保持稳定,都需要通过控制训练与仿真反复验证。这些试错代价高昂,其中的经验却往往没有被整理成可供后续设计复用的知识。
密歇根大学安娜堡分校研究团队提出了AI机器人协同设计智能体Auto-Robotist,相关工作已获EMNLP 2026主会录用并开源。