<h1 align="center"> <a href="https://evggt.github.io/">Improving Robotic Manipulation with Efficient Geometry-Aware Vision Encoder<br></a> </h1> <div align="center">

   

</div>

Installation

Create a conda environment

$ conda create -n eVGGT python=3.10 -y
$ conda activate eVGGT

To install basic environments and CuRobo:

$ bash script/_install.sh

Download assets for RoboTwin:

$ bash script/_download_assets.sh

Install required packages:

$ cd policy/VGGT/vggt/
$ pip install -r requirements.txt
$ pip install -e . # Install VGGT dependency
cd ..
$ pip install -e . # Install diffusion_policy depenedency
$ cd ../..

For weights of eVGGT used for training policy, please download at this link. Place it at:

policy/VGGT/checkpoints/distillation

For more details of RoboTwin, follow RoboTwin 2.0 Document (Usage - Install & Download).

Training & Evaluation

Step 1: Collect data:

$ cd ../..
$ bash collect_data.sh ${task_name} ${task_config} ${gpu_id}
# Example: bash collect_data.sh beat_block_hammer demo_randomized 0

• Step 2: Process data:

$ cd policy/VGGT
$ bash process_data.sh ${task_name} ${task_config} ${expert_data_num}
# bash process_data.sh beat_block_hammer demo_randomized 50

• Step 3: Train policy:

$ bash train.sh ${task_name} ${task_config} ${expert_data_num} ${seed} ${action_dim} ${gpu_id}
# bash train.sh beat_block_hammer demo_randomized 50 0 14 0
# For `aloha-agilex` embodiment, the action_dim is 14

• Step 4: Eval policy:

$ bash eval.sh ${task_name} ${task_config} ${ckpt_setting} ${expert_data_num} ${seed} ${gpu_id}
# bash eval.sh beat_block_hammer demo_randomized demo_randomized 50 0 0
# This command trains the policy using the `demo_randomized` setting ($ckpt_setting)
# and evaluates it using the same `demo_randomized` setting ($task_config).
#
# To evaluate a policy trained on the `demo_randomized` setting and tested on the `demo_clean` setting, run:
# bash eval.sh beat_block_hammer demo_clean demo_randomized 50 0 0

For other policies, please refer to RoboTwin 2.0 Document (Usage).

<!-- # Citations If you find this work useful, please consider citing: <b>RoboTwin 2.0</b>: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation ``` @article{chen2025robotwin, title={RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation}, author={Chen, Tianxing and Chen, Zanxin and Chen, Baijun and Cai, Zijian and Liu, Yibin and Liang, Qiwei and Li, Zixuan and Lin, Xianliang and Ge, Yiheng and Gu, Zhenyu and others}, journal={arXiv preprint arXiv:2506.18088}, year={2025} } ``` -->

Acknowledgement

Thanks to Tianxing Chen et al. for their amazing RoboTwin platform.