CAS Team's Maxwell Embodied AI Model Tops MetaWorld Benchmark with Record Scores

A team led by researchers Wu Bin and Wang Shaojiang at the Institute of Industrial AI, Chinese Academy of Sciences, has announced that its self-developed embodied intelligence large model, Maxwell, achieved the top position in the MetaWorld benchmark comparison. Its average success rate across four difficulty levels surpassed models listed on the Evo-SOTA leaderboard, with an overall success rate of 91.90% and a Hard-mode success rate of 95.67%, setting a new record for the leaderboard.

MetaWorld is an important simulation benchmark for robot multi-task manipulation research, covering 50 tasks including grasping, moving, door opening, and drawer operation. It evaluates robots' execution capabilities across different manipulation objects and task requirements.

Lightweight design and multi-task capability are notable features of Maxwell. The model has 1B parameters, supports single-GPU inference, and is designed for edge deployment.

In addition to MetaWorld, the team also evaluated the model in the LIBERO simulation environment, covering 4 task suites, 40 tasks, and 1,693 trajectories. The model achieved a success rate of up to 99.1% on some tasks.