Embodied First Exam Tests Coding Agents on Simulated Robots
Researchers from eight universities, including Northwestern University and Carnegie Mellon University, proposed Embodied First Exam (EFE), a unified benchmark that connects coding agents to simulated robots. EFE contains 1,156 tasks and 2,288 tests across 35 task suites, 23 simulators and 47 robot embodiments. On the 456-task core set in standard mode, Claude Fable 5.2 solved 50% of tasks, followed by GPT-6 Astra at 34% and GPT-6.1 Sol at 32%.
EFE provides a substantial common testbed for measuring whether coding agents can transfer from software tasks to robot control, while its privileged, standard and realtime modes expose the effect of state access and resets on performance.
新智元 2026-11-19 12:46 北京
新智元报道
编程智能体已经能在数字世界完成复杂任务。
借助编程接口,它们也可以指挥物理世界中的机器人。
但机器人任务和软件任务有本质不同:智能体只能通过传感器了解环境,动作做出去往往无法撤回,环境也不会停下来等它思考。
#Embodied First Exam (EFE)#Northwestern University#Carnegie Mellon University#Claude Fable 5.2#Claude Opus 5.5#GPT-6 Astra#GPT-6.1 Sol#Gemini 4 Argon#Kimi 3.1 Pro