How does LLM control physical robots? Well, the gap between coagulation code and the robot is quite large. Obviously one of the next big things is the destruction of physical robots. First simulation Kangrui Wang, Qineng Wang, Chang Chang, Pingyue Jo, Chi Gui, Feishi Wang, Canyu Chen, Wenli Xiaoyu, Xiaoyu, Tongheangheang, Linjie Li, Zihan Wang, Minyeong Kim, Liu Liu Yu, Joo Hou, Jo, Han, Leo Ping, Ja, Zjo, Zjo, Z. What they did: First evaluation (EFE) is a score of 1,156 tasks on 47 simulated robots. In it, ordinary coding agents, without changes, direct problem actions, write controllers or train politicians. On complex tasks, the best proven agent (GPT-6 Astra in OpenAI CLI Codex) drops from 77% success when accessing simulator mode and infinite zeros to 30% with only sensors and limited zeros. This is mainly due to the fact that the exact steps require repeated attempts. Why it's important: In order to transfer agents to real iron, in addition to strong models, thoughtful execution systems and the safe restoration of the state are needed. However, all test estimates were obtained from individual simulation attempts. Go to the simulation here: