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 π0.5 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.
Figures & tables
Figure 1: Overview of Meituan-Robotics-0. The framework integrates large-scale multimodal data, three-stage VLA training, an event-driven embodied-agent system, and real-robot deployment. The RoboDojo panel orders the displayed VLA models in ascending order of average success rate.
Capability family
Training setting
Avg. Len.
Task 1
Task 2
Task 3
Task 4
Task 5
Baseline
–
3.832
91.0%
83.0%
76.2%
70.2%
62.8%
Spatial localization
+ 2D Grounding
3.941
91.3%
83.7%
78.1%
73.2%
67.8%
+ 2D Pointing
3.806
90.8%
81.8%
74.8%
69.9%
63.3%
+ 3D Grounding
4.037
92.4%
85.3%
79.6%
75.7%
70.7%
Action-related understanding
+ 2D Target-Affordance Grounding
4.102
93.3%
87.9%
82.5%
76.4%
70.1%
+ 3D Target-Affordance Grounding
4.098
93.7%
89.9%
81.2%
73.9%
71.1%
Table 1: Ablation of embodied vision–language supervision in the CALVIN ABC → D setting. Bold indicates the highest Avg. Len. within each capability family.
Figure 2: Vision-language dataset. The corpus contains approximately 147 million samples before training-time resampling, comprising 78% general and 22% embodied vision-language data. Representative examples are shown on the right. Table 3 reports the effective pre-training mixture after resampling.
Figure 3: Composition of the robot corpus. The 15,060-hour corpus spans 12 embodiments and comprises RoboDesk, a curated collection of open-source real-robot trajectories totaling 3,235 hours (21.5%); 7,530 hours (50.0%) of licensed real-robot data; and 4,295 hours (28.5%) of open simulation data.
Data source
Hours
Share
RoboDesk (real-world): 3,235 h (21.5%)
AgiBotWorld-Beta ( AgiBot-World-Contributors and others, 2025 )
2,174
14.4%
Galaxea Open-World ( Jiang et al., 2025 )
333
2.2%
RoboCOIN ( Wu and others, 2025 )
311
2.1%
LET ( LejuRobotics, 2025 )
172
1.1%
DROID ( Khazatsky et al., 2024 )
164
1.1%
Table 2: Open-source datasets retained after curation.
Figure 4: Robot trajectory curation and alignment pipeline. S1–S3 remove static boundaries, kinematic inconsistencies, and video defects; A1 canonicalizes actions across embodiments, and A2 aligns multi-granularity task annotations with the retained trajectories.
Table 3: Pre-training data mixture. Each optimization step draws one batch from the VLA action pool and one batch from the VLM co-training pool, with loss weights λvla=1.0 and λvlm=0.1 . VLM entries are organized by training group and report effective sample counts after resampling, including repeated samples. Within-pool ratios are computed from VLA frame counts and VLM effective sample counts, respectively.
Figure 6: Rollback data collection. Dots mark the rollback anchors t1 and t2 . Wavy blue branches (1) and (2) are replaced policy segments ( τ− , I=0 ); solid orange HIL segments are their human replacements ( τ+ , I=1 ), with corrected execution continuing toward (3). These conditions are used by CFG-DAgger. Gray segments denote other autonomous execution.
Figure 7: Asynchronous action-chunk merging. Predictions made stale during inference are removed from the chunk; the aligned overlap is blended with the residual buffer, and the suffix is appended for continuous execution.
Figure 8: Event-driven embodied-agent framework. The asynchronous agent loop converts visual observations into structured captions, consolidates them into events, and triggers the Planner when task-state reassessment is required; the active subtask conditions the high-frequency VLA control loop.
Figure 9: Real-robot evaluation tasks. Representative stages of (a) letter-block placement, (b) garment folding, and (c) Ethernet insertion.
Task
Demo
HIL / Rollback
Total
Training Samples
Letter-block placement
500
200
700
∼ 900K
Garment folding
400
400
800
∼ 1.1M
Ethernet insertion
200
450
650
∼ 300K
Table 6: Real-world post-training data. Demo counts teleoperated demonstrations; HIL / Rollback counts correction episodes collected with each task’s mechanism. Training Samples counts valid 50-step action chunks after preprocessing. The Ethernet row applies to the HIL scaling and Rollback comparisons; the MR0 vs. π0.5 comparison uses over 1,400 Ethernet episodes.
Figure 10: Real-robot SR and PS of π0.5 and MR0 after SFT and HG-DAgger post-training. PS is displayed as a percentage in the figures.
Figure 11: Ethernet insertion performance across Rollback collection rounds. The demonstration set is fixed. Each collection round adds the same number of Rollback episodes. Round 0 is a shared SFT reference.
Figure 12: Effect of accumulating HIL data on the three tasks. The initial point uses demonstrations only. For each task, each subsequent round adds the same number of HIL episodes. Each policy is trained independently from the same mid-training checkpoint.
Figure 13: Event memory and cross-task retrieval. While screwdriver placement remains active, a person moves the previously placed key from the bottom drawer to the middle drawer. The stored event is later retrieved to locate the key for a new instruction.
Figure 14: Subtask waiting and condition recovery. Removing the key makes the active placement subtask temporarily infeasible. The agent retains the pending objective and resumes execution after the key returns.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Quantity or option
Value
Role
Strategy
Temporal chunk-wise smoothing
Background prediction with latency-aware buffer updates.
fc
30Hz
Command-consumption and publication rate of the control loop.
fπ
3Hz
Target background prediction rate.
H
50
Action horizon returned by the policy server.
Klat
4
Maximum number of leading commands removed from an incoming chunk.
mmin
8
Minimum old-reference length used for short-buffer padding.
Appendix
Table 7: Configuration of the asynchronous executor.
Parameter
Setting
Value support [Vmin,Vmax]
[−1,0]
Number of value atoms B
201
Value learning steps / batch size
6,000 / 256
Value peak learning rate
5×10−5
Value warm-up steps
500
Value final learning rate
1×10−6
Appendix
Table 8: RECAP value learning, advantage labeling, and policy-training hyperparameters.