README.md

---

datasets:
  - lerobot/pusht
library_name: lerobot
license: apache-2.0
model_name: diffusion
pipeline_tag: robotics
tags:
  - lerobot
  - robotics
  - diffusion
  - pusht
  - imitation-learning
  - benchmark
---



# 🦾 Diffusion Policy for Push-T (200k Steps)

[![LeRobot](https://img.shields.io/badge/Library-LeRobot-yellow)](https://github.com/huggingface/lerobot)
[![Task](https://img.shields.io/badge/Task-Push--T-blue)](https://huggingface.co/datasets/lerobot/pusht)
[![UESTC](https://img.shields.io/badge/Author-UESTC_Graduate-red)](https://www.uestc.edu.cn/)
[![License](https://img.shields.io/badge/License-Apache_2.0-green)](https://www.apache.org/licenses/LICENSE-2.0)

> **Summary:** This model demonstrates the capabilities of **Diffusion Policy** on the precision-demanding **Push-T** task. It was trained using the [LeRobot](https://github.com/huggingface/lerobot) framework as part of a thesis research project benchmarking Imitation Learning algorithms.
- **🧩 Task**: Push-T (Simulated)
- **🧠 Algorithm**: [Diffusion Policy](https://huggingface.co/papers/2303.04137) (DDPM)
- **🔄 Training Steps**: 200,000 (Fine-tuned via Resume)
- **🎓 Author**: Graduate Student, **UESTC** (University of Electronic Science and Technology of China)
---

## 🔬 Benchmark Results (vs ACT)

Compared to the ACT baseline (which achieved **0%** success rate in our controlled experiments), this Diffusion Policy model demonstrates significantly better control precision and trajectory stability.

### 📊 Evaluation Metrics (50 Episodes)

| Metric | Value | Comparison to ACT Baseline | Status |
| :--- | :---: | :--- | :---: |
| **Success Rate** | **14.0%** | **Significant Improvement** (ACT: 0%) | 🏆 |
| **Avg Max Reward** | **0.81** | **+58% Higher Precision** (ACT: ~0.51) | 📈 |
| **Avg Sum Reward** | **130.46** | **+147% More Stable** (ACT: ~52.7) | ✅ |

> **Note:** The Push-T environment requires **>95% target coverage** for success. An average max reward of `0.81` indicates the policy consistently moves the block very close to the target position, proving strong manipulation capabilities despite the strict success threshold.
 
---

## ⚙️ Model Details

| Parameter | Description |
| :--- | :--- |
| **Architecture** | ResNet18 (Vision Backbone) + U-Net (Diffusion Head) |
| **Prediction Horizon** | 16 steps |
| **Observation History** | 2 steps |
| **Action Steps** | 8 steps |

- **Training Strategy**: 
  - Phase 1: Initial training (100,000 steps) -> Model: `Lemon-03/DP_PushT_test`
  - Phase 2: Resume/Fine-tuning (+100,000 steps) -> Model: `Lemon-03/DP_PushT_test_Resume`
  - **Total**: 200,000 steps
---

## 🔧 Training Configuration (Reference)

For reproducibility, here are the key parameters used during the training session:

  - **Batch Size**: 64
  - **Optimizer**: AdamW (`lr=1e-4`)
  - **Scheduler**: Cosine with warmup
  - **Vision**: ResNet18 with random crop (84x84)
  - **Precision**: Mixed Precision (AMP) enabled

#### Original Training Command (My Resume Mode)

```bash
python -m lerobot.scripts.lerobot_train \
  --policy.type diffusion \
  --env.type pusht \
  --dataset.repo_id lerobot/pusht \
  --wandb.enable true \
  --eval.batch_size 8 \
  --job_name DP_PushT_Resume \
  --policy.repo_id Lemon-03/DP_PushT_test_Resume \
  --policy.pretrained_path outputs/train/2025-12-02/14-33-35_DP_PushT/checkpoints/last/pretrained_model \
  --steps 100000
```

---

## 🚀 Evaluate (My Evaluation Mode)

Run the following command in your terminal to evaluate the model for 50 episodes and save the visualization videos:

```bash
python -m lerobot.scripts.lerobot_eval \
  --policy.type diffusion \
  --policy.pretrained_path outputs/train/2025-12-04/14-47-37_DP_PushT_Resume/checkpoints/last/pretrained_model \
  --eval.n_episodes 50 \
  --eval.batch_size 10 \
  --env.type pusht \
  --env.task PushT-v0
```

To evaluate this model locally, run the following command:

```bash
python -m lerobot.scripts.lerobot_eval \
  --policy.type diffusion \
  --policy.pretrained_path Lemon-03/DP_PushT_test_Resume \
  --eval.n_episodes 50 \
  --eval.batch_size 10 \
  --env.type pusht \
  --env.task PushT-v0
```

-----