Models
Meituan’s robotics team posts Meituan-Robotics-0 (MR0), a vision-language-action model for desktop manipulation trained on 15,060 hours of robot trajectories across 12 embodiments; the benchmark results in the arXiv paper are self-reported
Meituan’s robotics team posted MR0, a vision-language-action model for desktop manipulation, on arXiv with 15,060 hours of robot data. Self-reported.
Meituan’s robotics team posted Meituan-Robotics-0 (MR0) on arXiv on November 5, 2026. It describes MR0 as a vision-language-action foundation model and embodied-agent system for desktop manipulation with a stationary base. The paper says it combines 147 million vision-language samples with 15,060 hours of curated robot trajectories spanning 12 embodiments. The results come from the authors’ own evaluation and the paper has not been peer reviewed.
The essentials
- The data: of the 15,060 hours, about 3,235 hours (21.5%) come from RoboDesk, a collection of open-source real-robot trajectories, 7,530 hours (50.0%) are licensed real-robot data and 4,295 hours (28.5%) are open simulation data. The paper also says about 13.6k hours of embodied data were used in training.
- The results: the abstract reports 93.6% success on RoboTwin 2.0 and fifth place in the RoboDojo VLA track. It also says that under matched HG-DAgger post-training MR0 beats π0.5 on average success and progress across three real-robot tasks.
- The method: a unified end-effector interface supports different robot bodies, and asynchronous action-chunk merging keeps motion smooth when inference latency varies. The GitHub page describes a first stage that learns discrete action tokens and a second that moves to continuous control with a flow-matching model.
- The tasks: the repository mentions real-world demonstrations of letter block placement, garment folding and network cable insertion. It says MR0 builds on the StarVLA training and deployment stack.
What is established and what is not
Established: the paper exists on arXiv, with 37 authors listed, and states the data composition and benchmark numbers above. Not established: independent replication; how MR0 performs outside the tested desktop settings; whether the model weights are released, which we did not confirm from the repository text we saw. “Desktop” here means a stationary arm at a table, not a mobile or humanoid robot.
Why it matters
Large robot-action datasets are becoming a main input for general robot models, and a company better known for delivery is publishing its own recipe and numbers. A fifth-place ranking and a three-task comparison with π0.5 are narrow evidence. Not investment advice.
What is not established
- Independent verification of the 93.6% and ranking results: none found.
- Public release of MR0 weights: not confirmed in the sources read.
- Performance on mobile or humanoid robots: outside the paper’s scope.
- Peer review: the paper is a preprint.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Submitted | November 5, 2026 | arXiv |
| Training data | 147M vision-language samples, 15,060 hours, 12 embodiments | arXiv |
| Data split | 3,235 h RoboDesk, 7,530 h licensed real, 4,295 h simulation | arXiv HTML |
| Reported results | 93.6% on RoboTwin 2.0; fifth in RoboDojo VLA | arXiv |
| Code | Meituan-Robotics-0 repository, built on StarVLA | GitHub |
FAQ
What does MR0 do?
It is a model that turns camera images and language instructions into actions for a robot arm working at a desk.
Are the results independent?
No. They are reported by Meituan’s team in the paper.
Is it a humanoid model?
No. The paper targets stationary-base desktop manipulation.
Method and limits
Sources checked on November 9, 2026: the arXiv abstract page, the arXiv HTML version and the GitHub repository, read as search extracts, not in full. We did not run the model or review the weights. Not investment advice.
Sources
- arXiv: Meituan-Robotics-0: A Vision-Language-Action Foundation Model for Desktop Manipulation (abstract) (verified 2026-11-08)
- arXiv HTML: Meituan-Robotics-0: A Vision-Language-Action Foundation Model for Desktop Manipulation (HTML) (verified 2026-11-08)
- GitHub: Meituan-Robotics/Meituan-Robotics-0 (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy