Overview

This repo contains code and instructions to support the use of Xtrainer robot:

• Collecting Demonstration Data
• Training ACT Policy
• Training DP Policy
• Deploying Policies on Hardware

Installation

• System requirements: Ubuntu 20.04, Python 3.8
• CUDA, cuDNN and Anaconda is required for GPU acceleration.
• Setup virtual environment and install dependencies:

  conda create -n xtrainer_clover python==3.8.10
  conda activate xtrainer_clover
  pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118
  pip install -r requirements.txt
  pip install -e third_party/DynamixelSDK/python
  pip install -e ModelTrain/detr
  pip install -e robomimic-r2d2
  sudo apt install libxcb-cursor0
  

Collecting Demonstration Data

• Initialize the robot and camera nodes:

  python 1_find_port.py # Find the port of the robot
  python scripts > 3_just_buttonA.py # adjust the position between leader robot and follower robot
  python 2_get_offset.py # Get the offset of the robot
  python launch_nodes.py # Launch the robot and camera nodes
  

• Collect demonstration data for ACT training:

   python run_control.py --agent_name act
  

• Collect demonstration data for DP training:

   python run_control.py --agent_name dp
  

Training and Deploying Policies

ACT Training

1. Run python 4_collect2train_data.py to process the collected data. Note to change the dataset_name and task_name in the script.
2. Change the following arguments in ModelTrain/constants.py:
   • DATA_DIR: the path to the processed data.
   • TASK_CONFIGS: the task configuration for the model.
3. Run python ModelTrain/model_train.py to train the model, some important Training Arguments:
   • --ckpt_dir : the path to save the model.
   • --task_name : the task name for the model.
   • --batch_size: the batch size for the model.

DP Training

1. Run python ModelTrain/dp/split_data.py to process the collected data.
2. Run python ModelTrain/dp/pipeline.py to train the model. Some important Training Arguments:
   • --data_path is the splitted trajectory folder, which is the output_path + data_name. (data_name should not include suffix like _train or _train_10)
   • --model_save_path is the path to save the model

Deployment

Run python experiments/run_inference.py to deploy the trained ACT or DP models. Important Arguments:

• --agent_name: the name of the agent. set to act for ACT policy and dp for DP policy.
• --act_ckpt_dir: the path to the trained ACT model.
• --dp_ckpt_dir: the path to the trained DP model.
• --episode_len: the length of the episode. The user can tailor the length of the action sequence for an autonomous reasoning run by modifying this parameter, e.g., when the the robot has not finished the target action when the autonomous reasoning is running, the step length can be increased; if the robot has some extra actions after the target action, the step length can be decreased.

Acknowledgement

This project was developed with help from the following codebases.

• diffusion_policy
• bidexdiffuser

The Team

BiDexDiffuser is developed and maintained by the CLOVER Lab (Collaborative and Versatile Robots Laboratory), CUHK.