diffusion_libero_object

Diffusion Policy — LIBERO single-task (book → caddy)

Diffusion Policy trained with LeRobot on one LIBERO task:

pick up the book and place it in the back compartment of the caddy

Trained from scratch on a laptop GPU (RTX 4050, 6 GB VRAM).

Evaluation

Evaluated in the LIBERO simulator (libero_10, task 5) — the same task the policy was trained on. All 10 rollouts use LIBERO's canonical initial states with randomised object poses.

Per-episode outcomes (1 = success): [1, 0, 1, 1, 0, 1, 1, 0, 1, 0]

Reproduce:

lerobot-eval \
  --policy.path=anuragbhandari-eng/diffusion_libero_object \
  --env.type=libero --env.task=libero_10 --env.task_ids="[5]" \
  --env.observation_height=256 --env.observation_width=256 \
  --eval.n_episodes=10 --eval.batch_size=1 --env.max_parallel_tasks=1 \
  --output_dir=eval_out

---

Model Details

• License: apache-2.0
• Robot type: panda (Franka)
• Cameras: agentview (image) + wrist (image2)

Inputs & Outputs

Inputs

Outputs

Training Dataset

• Repository: HuggingFaceVLA/libero
• Task: pick up the book and place it in the back compartment of the caddy
• Episodes used: 19 (episodes 27,28,47,55,61,64,81,103,104,109,111,127,133,136,141,147,154,158,159)
• Frames: 3 609
• Frame rate: 10.0 FPS

Training Configuration

---

How to Reproduce Training

pip install -e ".[libero]" --no-build-isolation
export MUJOCO_GL=egl
lerobot-train \
  --policy.type=diffusion \
  --dataset.repo_id=HuggingFaceVLA/libero \
  --dataset.episodes="[27,28,47,55,61,64,81,103,104,109,111,127,133,136,141,147,154,158,159]" \
  --batch_size=8 --steps=80000 \
  --policy.device=cuda \
  --policy.push_to_hub=true \
  --policy.repo_id=anuragbhandari-eng/diffusion_libero_object \
  --save_freq=5000

---

Citation

@misc{cadene2024lerobot,
    author = {Cadene, Remi and others},
    title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
    howpublished = "\url{https://github.com/huggingface/lerobot}",
    year = {2024}
}