Lemon-03/DP_PushT_test

🦾 Diffusion Policy for Push-T (Phase 1: 100k Steps)

Summary: This model represents the initial training phase (0 - 100k steps) of a Diffusion Policy on the Push-T task. It serves as the pre-trained foundation for further fine-tuning. While it demonstrates strong trajectory learning capabilities, it has not yet fully converged to high success rates.

• 🧩 Task: Push-T (Simulated)
• 🧠 Algorithm: Diffusion Policy (DDPM)
• 🔄 Training Steps: 100,000 (Initial Phase)
• 🎓 Author: Graduate Student, UESTC (University of Electronic Science and Technology of China)

---

⚠️ Note on Performance & Fine-tuning

This checkpoint represents the intermediate state of our research. While it achieves high movement precision (Avg Max Reward: 0.71), the strict success threshold of the Push-T task results in a lower success rate at this stage.

🚀 Upgrade Available:

We performed Resume Training (Fine-tuning) based on this checkpoint for another 100k steps, achieving significantly better results. 👉 Check out the final model here: Lemon-03/DP_PushT_test_Resume

---

🔬 Benchmark Results (Phase 1)

Evaluated on 50 episodes in the Push-T environment using LeRobot.

Metric	Value	Status
Success Rate	4.0%	🚧 (Under-trained)
Avg Max Reward	0.71	📈 (High Precision)
Avg Sum Reward	115.03	✅ (Good Trajectory)

Analysis: The model has successfully learned the multimodal distribution of the demonstration data and can push the T-block close to the target (Reward 0.71). However, it lacks the final fine-grained adjustment capabilities required for the >95% overlap success criteria. This motivated the subsequent Phase 2 (Resume Training).

---

⚙️ 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 Configuration (Reference)

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

• Batch Size: 8 (Effective)
• Optimizer: AdamW (lr=1e-4)
• Scheduler: Cosine with warmup
• Vision: ResNet18 with random crop (84x84)

Original Training Command (My Training Mode)

python -m lerobot.scripts.lerobot_train \
  --policy.type diffusion \
  --env.type pusht \
  --dataset.repo_id lerobot/pusht \
  --wandb.enable true \
  --job_name DP_PushT \
  --policy.repo_id Lemon-03/DP_PushT_test \
  --eval.batch_size 8

---

🚀 Evaluate (My Evaluation Mode)

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

python -m lerobot.scripts.lerobot_eval \
  --policy.type diffusion \
  --policy.pretrained_path outputs/train/2025-12-02/14-33-35_DP_PushT/checkpoints/last/pretrained_model \
  --eval.n_episodes 50 \
  --eval.batch_size 10 \
  --env.type pusht \
  --env.task PushT-v0

You can evaluate this checkpoint to reproduce the Phase 1 results:

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