Top of the Meta-World ranking
Maxwell scored 91.9 on Meta-World, ahead of FabriVLA from Shenzhen-based Youibot with 90 and SUREFlow from South Korea’s Kyungpook National University with 88.3. The ranking draws leading teams worldwide, including Physical Intelligence, Google DeepMind, Alibaba, Meituan, Carnegie Mellon University, MIT, the University of Cambridge and the Toyota Research Institute.
To complete a task in Meta-World, a model must understand the spatial relationship between objects and targets, carry out a sequence of actions such as approaching, grasping and moving, and adjust based on contact.
Key results:
- Meta-World: Maxwell 91.9, FabriVLA 90, SUREFlow 88.3.
- More than 200 tasks performed without additional fine-tuning.
- 1% success rate in the LIBERO environment.
- 1 billion parameters, running on a single card.
From thinking to acting
In simulated demonstrations, Maxwell could move a nut onto a target post, open cabinet doors and drawers using a handle, and carry objects between containers. In LIBERO, which emphasises continuous operations, it linked multiple steps such as turning on a stove and putting a kettle on it, placing a cup in a microwave and putting a book in a storage box.

















