syscv / soccer-player

Dribbling Benchmark Environments

This repository provides an implementation of the paper:

DexDribbler: Learning Dexterous Soccer Manipulation via Dynamic Supervision

[Paper]|[Website]|[Video]|

Our work is greatly inspired by DribbleBot and built upon IsaacGymEnvs.

Installation

Isaac Gym

Download the Isaac Gym Preview 4 release from the website, then follow the installation instructions in the documentation. We highly recommend using a conda environment to simplify set up.

Ensure that Isaac Gym works on your system by running one of the examples from the python/examples directory, like joint_monkey.py. Follow troubleshooting steps described in the Isaac Gym Preview 4 install instructions if you have any trouble running the samples.

Gym Envs

Once Isaac Gym is installed and samples work within your current python environment, install this repo:

pip install -e .

To ensure the environments work properly, please execute the following command:

python train.py

Ants should learn to walk forward as fast as possible.

Load trained models for evaluation

Absolute Tracking Error

python ./script/box_plot.py 

data for box plot has already been generated, you can run the experiment script to regenerate.

bash ./script/test_3x5_dribble.sh

Trajectory Following

python ./script/traj_plot.py

data for trajectory plot has already been generated, you can run the experiment script to regenerate.

bash ./script/test_2x2_traj.sh

Train your own Dribbler

DexDribbler(Ours)

basically DribbleBot + estimate more context parameter + dynamic feedback supervision

python train.py task=Go1Dribble train=Go1DribblePPOsea seed=42 

DribbleBot+

dynamic feedback supervision disabled, only DribbleBot + estimate more context parameter

python train.py task=Go1Dribble train=Go1DribblePPOsea seed=42 ~task.env.rewards.rewardScales.tracking_lin_vel_PID ~task.env.rewards.rewardScales.raibert_heuristic_PID

DribbleBot (Baseline)

original DribbleBot, but extend the lower bound of ball-terrain-drag from 0.1 to-0.1

python train.py task=Go1Dribble train=Go1DribblePPOsea seed=42 ~task.env.rewards.rewardScales.tracking_lin_vel_PID ~task.env.rewards.rewardScales.raibert_heuristic_PID ~task.env.priviledgeStates.ball_states_v_1 ~task.env.priviledgeStates.ball_states_p_1 ~task.env.priviledgeStates.ball_states_v_2 ~task.env.priviledgeStates.ball_states_p_2 ~task.env.priviledgeStates.dof_stiff ~task.env.priviledgeStates.dof_damp ~task.env.priviledgeStates.dof_calib ~task.env.priviledgeStates.payload ~task.env.priviledgeStates.com ~task.env.priviledgeStates.friction ~task.env.priviledgeStates.restitution ~task.env.priviledgeStates.ball_mass ~task.env.priviledgeStates.ball_restitution

Reproducibility

the random seeds we used in the experiment part of the paper are [42,3,13,17,69,106].

Dribbling under other leg configuration

By simply adapting the action dimension and specifying the foot index number, our training pipeline and the original DribbleBot approach are both naturally compatible for ball dribbling tasks across different legged robot configurations.

you can try:

python train.py task=NaoDribble train=Go1DribblePPOsea

or

python train.py task=CaDribble train=Go1DribblePPOsea

But they are not verified in real world.

Citing

Please cite this work as:

@misc{hu2024dexdribbler,
      title={DexDribbler: Learning Dexterous Soccer Manipulation via Dynamic Supervision}, 
      author={Yutong Hu and Kehan Wen and Fisher Yu},
      year={2024},
      eprint={2403.14300},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}