Meituan-Robotics-0: A Vision-Language-Action Foundation Model for Desktop Manipulation

Meituan Robotics Team, Caicheng Wang, Chen Zhao, Chengpeng Jin, Duo Xu, Fengjiao Chen, Guanglu Wan, Heqing Yang, Jia Wang, Jiacong Zhou, Jie Hu, Jiyuan Ru, Lingfeng Zhang, Linjiang Huang, Manyuan Zhang, Ningchu Liu, Shijie Zhou, Si Liu, Shuo Li, Tianyi Zhang, Xi Zhang, Xin Chen, Xuezhi Cao, Xunliang Cai, Yanbin Zeng, Yang Yang, Yang Yi, Yanli Tan, Yifu Xu, Yimeng Jia, Yuchen Tang, Yuchen Xie, Yuxiang Zheng, Zerwei Chng, Zilong Huang, Zipeng Ye, Ziyuan Qin, Ziqin Wang

November 2026

Real-robot garment folding with the bimanual arms.

TL;DR A tabletop vision-language-action model trained on 147M vision-language samples and 15,060 hours of robot data from 12 embodiments, then post-trained on a real bimanual robot with human corrections. It scores 93.6% on RoboTwin 2.0, and its rewind-and-correct data collection reaches 70% success on real Ethernet insertion.

What they do

MR0 is a vision-language-action (VLA) policy plus an agent layer for stationary tabletop manipulation. The policy is trained in three stages on 147M vision-language samples and 15,060 hours of curated trajectories from 12 embodiments, then adapted on a bimanual AgileX PiPER with human corrections. Because the policy only reacts to the current observation, a separate event-driven agent keeps track of what happened across subtasks.

How it works

One action space. Trajectories are trimmed and filtered for kinematic and video defects. Their actions are converted to end-effector position, 6D rotation and gripper per arm in a shoulder-centered frame, so single-arm and bimanual robots share one 20-dimensional action.
Three-stage training. A Qwen3.5-4B backbone is first pre-trained on FAST action tokens alongside VQA data. An MMDiT flow-matching head then takes over for continuous control, and the policy is finally fine-tuned on real demonstrations and corrections.
Rollback and CFG-DAgger. Near a failure the operator rewinds the robot to an earlier state and demonstrates a fix. The replaced policy segment becomes a negative label and the human fix a positive one. At inference, classifier-free guidance steers away from the negative, so no value function is needed.
Agent. A Monitor (Qwen3.5-35B-A3B, 2 Hz) captions short windows, a Memory Manager stores them as events, and a Planner runs only when an event changes progress or feasibility.

Results

RoboTwin 2.0 (50 bimanual simulation tasks): 93.6% average success (93.2% clean, 93.9% randomized), 8.0 points above ABot-M0.
RoboDojo: fifth by success rate in the VLA track (14.63%) with about 13.6k hours of robot data. Open-vocabulary instructions reach only 3.38%.
Real robot (letter blocks, garment folding, Ethernet insertion; 30 trials each): neither MR0 nor π0.5 folds a garment after SFT alone. After HG-DAgger, MR0 ties π0.5 on letter-block success and gains most on Ethernet insertion.
Ethernet insertion with the same Rollback data: CFG-DAgger reaches 70% success and 0.90 progress, against 60% and 0.67 for RECAP.

Limitations

The agent is shown in two qualitative deployments, not measured. Real-robot tests use one bimanual platform and three tasks. Unfamiliar instructions are weak; for example, on RoboDojo’s Align_blocks task the policy sometimes grasps the box instead of aligning it.

Videos

Real-robot Ethernet insertion: grasp the cable, hand it between arms and insert it into a port.

Agent demo where a new goal makes the planner recall the key's last seen drawer from event memory and replan.

Figures

From Figure 1: Overview of Meituan-Robotics-0: large-scale multimodal data, three-stage VLA training, an event-driven... — From Figure 1: Overview of Meituan-Robotics-0: large-scale multimodal data, three-stage VLA training, an event-driven embodied-agent system and real-robot deployment.

From Figure 8: Event-driven embodied-agent framework. The agent loop turns visual observations into structured captio... — From Figure 8: Event-driven embodied-agent framework. The agent loop turns visual observations into structured captions, consolidates them into events and triggers the Planner when task state needs reassessing; the active subtask conditions the high-frequency VLA control loop.

From Figure 10: Real-robot success rate and progress score of π0.5 and MR0 after SFT and after HG-DAgger post-training.

Abstract

We present Meituan-Robotics-0 (MR0), a vision-language-action foundation model and embodied-agent system for stationary-base desktop manipulation. MR0 combines 147 million vision-language samples with 15,060 hours of curated robot trajectories spanning 12 embodiments. A three-stage training recipe aligns multimodal representations with robot behavior through discrete action-token pre-training, transfers these priors to continuous cross-embodiment control through flow-matching mid-training, and adapts the policy through real-world post-training. We introduce Rollback, which pairs replaced policy segments with human corrections to provide positive and negative supervision for CFG-DAgger without value learning or advantage labeling. A unified end-effector interface supports heterogeneous embodiments, while asynchronous action-chunk merging maintains smooth execution under variable inference latency. An event-driven agent maintains persistent task events and reassesses plans when progress or feasibility changes, enabling cross-task retrieval and conditional task resumption. MR0 achieves 93.6% success on RoboTwin 2.0 and ranks fifth in the RoboDojo VLA track. Under matched HG-DAgger post-training, MR0 achieves higher average success rate and progress score than $π_{0.5}$ across three real-robot tasks. On Ethernet insertion, CFG-DAgger trained with Rollback data reaches 70% success and a progress score of 0.90. Qualitative agent deployments demonstrate cross-task event retrieval and recovery from temporarily unsatisfied task conditions.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong