---
tags:
- lerobot
- diffusion-policy
- pusht
- simulation
---

# Diffusion Policy on PushT - results

Training and evaluation artifacts for the policy [`NLTuan/diffusion_pusht_long`](https://huggingface.co/NLTuan/diffusion_pusht_long).

## Final evaluation (PushT sim)

```json
{
  "avg_sum_reward": 104.10700966672937,
  "avg_max_reward": 0.8298583084048309,
  "pc_success": 34.0,
  "n_episodes": 50,
  "n_success": 17,
  "pc_success_ci95": [
    22.43694922250009,
    47.84617394976739
  ],
  "eval_s": 53.94045829772949,
  "eval_ep_s": 1.0788091659545898,
  "video_paths": [
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_0.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_1.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_2.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_3.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_4.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_5.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_6.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_7.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_8.mp4",
    "outputs/eval/diffusion_pusht_long/videos/pusht_0/eval_episode_9.mp4"
  ],
  "predicted_video_paths": []
}
```

## Contents

- `train_eval/` - policy rollout videos captured during periodic in-training evaluation.
- `final_eval/` - final standalone `lerobot-eval` videos and `eval_info.json`.
- `train_config.json` - exact training configuration.
- `train.log` - full training stdout.

## Reproduce

```bash
lerobot-train --policy.type=diffusion --policy.device=cuda --ema.enable=true \
  --env.type=pusht --dataset.repo_id=lerobot/pusht \
  --batch_size=64 --steps=20000 \
  --policy.push_to_hub=true --policy.repo_id=NLTuan/diffusion_pusht_long \
  --output_dir=outputs/train/diffusion_pusht

lerobot-eval --policy.path=NLTuan/diffusion_pusht_long --env.type=pusht \
  --eval.n_episodes=50 --eval.batch_size=25
```