Meituan's robotics team has described Meituan-Robotics-0 (MR0), a vision-language-action foundation model and agent system for robots that handle objects on a desk from a fixed base, in a paper on arXiv. It was trained on 147 million vision-language samples plus 15,060 hours of selected robot trajectories from 12 robot types, in three stages that end with post-training in the real world.
The team adds Rollback, which pairs replaced segments of the robot's actions with human corrections to give positive and negative examples, and an event-driven agent that tracks tasks and revises plans when progress stalls. MR0 reaches 93.6% success on the RoboTwin 2.0 benchmark and ranks fifth in the RoboDojo VLA track, the authors report.
Background
Meituan
Technology company
A Chinese company whose robotics team built the Meituan-Robotics-0 model.
Description from arXiv Robotics (cs.RO)
Meituan-Robotics-0
Robot foundation model
A vision-language-action foundation model and agent system for robots that manipulate objects on a desk.
Description from arXiv Robotics (cs.RO)
