SeedPolicy: Horizon Scaling via Self-Evolving Diffusion Policy for Robot Manipulation

This repository contains the pre-trained model checkpoints for the tasks highlighted in the paper SeedPolicy: Horizon Scaling via Self-Evolving Diffusion Policy for Robot Manipulation.

📄 Associated Paper & Links

• Hugging Face Paper Page: https://huggingface.co/papers/2603.05117
• GitHub Repository: https://github.com/Youqiang-Gui/SeedPolicy

💡 Overview

SeedPolicy introduces Self-Evolving Gated Attention (SEGA), a temporal module that maintains a time-evolving latent state via gated attention. This enables efficient recurrent updates that compress long-horizon observations into a fixed-size representation while filtering irrelevant temporal information. Integrating SEGA into Diffusion Policy (DP) resolves temporal modeling bottlenecks and enables scalable horizon extension for long-horizon robotic manipulation tasks.

🛠️ Usage

Detailed installation and data generation instructions are available in the official GitHub repository.

1. Train Policy

bash train.sh ${task_name} ${task_config} ${expert_data_num} ${seed} ${action_dim} ${gpu_id} ${config_name}

# Example:
# bash train.sh beat_block_hammer demo_clean 50 0 14 0 train_diffusion_transformer_hybrid_workspace

2. Eval Policy

bash eval.sh ${task_name} ${task_config} ${ckpt_setting} ${expert_data_num} ${seed} ${gpu_id} ${config_name} ${timestamp}

# Example 1: Standard Evaluation
# bash eval.sh beat_block_hammer demo_clean demo_clean 50 0 0 train_diffusion_transformer_hybrid_workspace "'20260106-143723'"

# Example 2: Generalization Evaluation
# To evaluate a policy trained on the `demo_clean` setting and tested on the `demo_randomized` setting, run:
# bash eval.sh beat_block_hammer demo_randomized demo_clean 50 0 0 train_diffusion_transformer_hybrid_workspace "'20260106-143723'"

The evaluation results, including videos, will be saved in the eval_result directory under the project root.

😺 Acknowledgements

Our code is generally built upon: Diffusion Policy and RoboTwin 2.0. Specifically, the implementation of our state update code references CUT3R and TTT3R.