Meituan-Robotics-0-RoboCasa365

Technical Report

MR0 policy adapted to RoboCasa365, with a Qwen3.5-4B backbone and an MMDiT action head. This is the supplied 200,000-step evaluation checkpoint.

Property	Value
Complete VLA weights	checkpoints/model.pt
Action dimensions / prediction horizon	12 / 50
State dimensions	20
RGB and state history	Four observations, offsets [-6, -4, -2, 0] at 20 Hz
Observation image size	640 × 320
Model configuration	config.full.yaml
Normalization statistics	dataset_statistics.json

checkpoints/model.pt contains both the Qwen backbone and action-head weights. base_vlm/ contains the matching architecture configuration, tokenizer, processor and chat template; it contains no separate Qwen weight file. The original steps_200000_pytorch_model.pt is a local compatibility link to the same checkpoint and is not duplicated here.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MeituanRobotics/Meituan-Robotics-0-RoboCasa365",
    local_dir="Meituan-Robotics-0-RoboCasa365",
)

Evaluation

Use the matching RoboCasa365 evaluation repository, tested at commit 380232d. Follow its setup instructions for the separate model and simulation environments. This policy requires its RGB/state history handling and Native12 action interface.

From the evaluation repository root, inspect the configuration before launching:

BUNDLE=/absolute/path/to/Meituan-Robotics-0-RoboCasa365
python scripts/run_eval.py \
  --checkpoint "$BUNDLE/checkpoints/model.pt" \
  --base-vlm "$BUNDLE/base_vlm" \
  --config "$BUNDLE/eval_8gpu.yaml" \
  --model-python /path/to/model/env/bin/python \
  --simulation-python /path/to/simulation/env/bin/python \
  --output /path/to/eval_output \
  --dry-run

Use eval_2gpu.yaml for two GPUs. After completing the evaluation repository's rendering preflight, remove --dry-run to run the evaluation. Both configurations select 50 episodes per task and execute 16 predicted actions per request.

The loader resolves configuration and statistics above the checkpoints/ directory, and locates base_vlm/ within the downloaded bundle. --base-vlm or EVAL_BASE_VLM can override the metadata directory.

The checkpoint and model/tokenizer metadata preserve the supplied bytes. Only the machine-specific path in config.full.yaml and the release documentation were made portable.

File verification

SHA256SUMS is the single SHA-256 checksum manifest for this release. It covers every repository file except the manifest itself, including weights, configuration, tokenizer/processor metadata, and documentation.

After downloading the complete repository, run from its root directory:

sha256sum -c SHA256SUMS