You can deploy this model and start earning money today!

Piersondong/
flow_matching_libero

Name: flow_matching_libero
Availability: InStock

HuggingFaceVLA/libero

2210.02747

Views

No views yet

Flow Matching for LIBERO

This repository contains the final LeRobot Flow Matching checkpoint trained on HuggingFaceVLA/libero. The policy learns a continuous velocity field from Gaussian noise to normalized action chunks and uses frozen CLIP task-text conditioning for the multi-task LIBERO dataset. It is the reproducible checkpoint associated with LeRobot PR #3359.

The implementation applies the straight Optimal Transport path from Flow Matching for Generative Modeling to robot action trajectories; it is a compact conditional baseline, not a reproduction of a robot-specific paper.

Training

Dataset: HuggingFaceVLA/libero
Steps: 100,000
Seed: 1000
Hardware: 4 x NVIDIA H20 96 GB
Precision: BF16 distributed training
Batch size: 64 per GPU, 256 effective
Data workers: 16 per rank
Checkpoints: every 10,000 steps; this repository contains the final 100,000-step checkpoint
Final logged training loss: 0.087

Exact training command:

bash

CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch \
  --multi_gpu \
  --num_machines=1 \
  --num_processes=4 \
  --mixed_precision=bf16 \
  --dynamo_backend=no \
  $(which lerobot-train) \
  --dataset.repo_id=HuggingFaceVLA/libero \
  --dataset.image_transforms.enable=true \
  --policy.type=flow_matching \
  --policy.text_encoder_name=openai/clip-vit-base-patch16 \
  --policy.freeze_text_encoder=true \
  --policy.use_group_norm=false \
  --policy.pretrained_backbone_weights=DEFAULT \
  --policy.device=cuda \
  --policy.push_to_hub=false \
  --output_dir=outputs/train/flow_matching_libero \
  --job_name=flow_matching_libero \
  --seed=1000 \
  --steps=100000 \
  --batch_size=64 \
  --num_workers=16 \
  --prefetch_factor=4 \
  --log_freq=100 \
  --save_checkpoint=true \
  --save_freq=10000 \
  --wandb.enable=false

LIBERO evaluation

The final checkpoint was evaluated once over all four standard suites with 10 tasks per suite and 10 episodes per task. This is 100 episodes per suite and 400 rollouts in total. Evaluation used batch size 1, one parallel task, seed 1000, and the LeRobot LIBERO environment's standard task horizons.

Suite	Tasks	Episodes per task	Episodes	Successes	Success rate
`libero_spatial`	10	10	100	13	13.0%
`libero_object`	10	10	100	4	4.0%
`libero_goal`	10	10	100	18	18.0%
`libero_10`	10	10	100	6	6.0%
Overall	40	10	400	41	10.25%

The overall rate is episode-weighted. Because every suite has 100 episodes, it is also the unweighted mean of the four suite rates. Machine-readable metrics are included in eval/results.json.

Public reproduction command:

bash

MUJOCO_GL=egl lerobot-eval \
  --policy.path=Piersondong/flow_matching_libero \
  --env.type=libero \
  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
  --eval.batch_size=1 \
  --eval.n_episodes=10 \
  --env.max_parallel_tasks=1 \
  --seed=1000 \
  --output_dir=outputs/eval/flow_matching_libero \
  --job_name=flow_matching_libero_standard_400

Exact successful evaluation invocation on the benchmark host (the extra variables only point to its non-interactive LIBERO configuration and environment-local EGL loader):

bash

CUDA_VISIBLE_DEVICES=0 \
MUJOCO_GL=egl \
PYOPENGL_PLATFORM=egl \
LIBERO_CONFIG_PATH=/data/dongyf/fastwam/config/libero \
LD_LIBRARY_PATH=/data/dongyf/fastwam/deps/libglvnd-jammy/root/usr/lib/x86_64-linux-gnu \
/data/dongyf/fastwam/envs/lerobot-libero/bin/lerobot-eval \
  --policy.path=/data/dongyf/fastwam/runs/flow_matching_libero_train_4gpu_20260803_093846/checkpoints/last/pretrained_model \
  --env.type=libero \
  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
  --eval.batch_size=1 \
  --eval.n_episodes=10 \
  --env.max_parallel_tasks=1 \
  --seed=1000 \
  --output_dir=/data/dongyf/fastwam/runs/flow_matching_libero_train_4gpu_20260803_093846/eval_standard_400 \
  --job_name=flow_matching_libero_standard_400

Files

model.safetensors: final policy parameters
config.json: policy architecture and feature configuration
policy_preprocessor.json and its normalizer weights
policy_postprocessor.json and its unnormalizer weights
train_config.json: recorded training configuration
eval/results.json: full overall, per-suite, and per-task metrics
eval/RESULTS.md: concise benchmark summary

Limitations

This research checkpoint has a 10.25% aggregate success rate under the stated protocol and is not suitable for safety-critical or real-world deployment. Results are specific to this training dataset, seed, checkpoint, simulator version, and evaluation configuration. No Arena evaluation is claimed.

Reference

bibtex

@inproceedings{lipman2023flow,
  title={Flow Matching for Generative Modeling},
  author={Yaron Lipman and Ricky T. Q. Chen and Heli Ben-Hamu and Maximilian Nickel and Matt Le},
  booktitle={The Eleventh International Conference on Learning Representations},
  year={2023},
  url={https://openreview.net/forum?id=PqvMRDCJT9t}
}