BAS

Official Implementation for Bridging Adaptivity and Safety: Learning Agile Collision-Free Locomotion Across Varied Physics.

L4DC 2025

Yichao Zhong, Chong Zhang, Tairan He, Guanya Shi

This codebase is under CC BY-NC 4.0 license, with inherited license in Legged Gym and RSL RL from ETH Zurich, Nikita Rudin and NVIDIA CORPORATION & AFFILIATES. You may not use the material for commercial purposes, e.g., to make demos to advertise your commercial products.

Please read through the whole README.md before cloning the repo.

Training in Simulation

Installation and Training Pipelines

Note: Before running our code, it's highly recommended to first play with RSL's Legged Gym version to get a basic understanding of the Isaac-LeggedGym-RslRL framework.

1. Create environment and install torch

   conda create -n bas python=3.8  # or use virtual environment/docker
   
   pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116  
   # used version during this work: torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2
   # for older cuda ver:
   pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
   

2. Install Isaac Gym preview 4 release https://developer.nvidia.com/isaac-gym

   unzip files to a folder, then install with pip:

   cd isaacgym/python && pip install -e .

   check it is correctly installed by playing:

   cd examples && python 1080_balls_of_solitude.py

3. Clone this codebase and install our rsl_rl in the training folder

   pip install -e rsl_rl

4. Install our legged_gym

   pip install -e legged_gym

   Ensure you have installed the following packages:

   • pip install numpy==1.21 (must < 1.24, >1.20)
   • pip install tensorboard
   • pip install setuptools==59.5.0

5. Try training.

   can use "--headless" to disable gui, press "v" to pause/resume gui play.

   for go1, in legged_gym/legged_gym,

   # BAS agile policy along with explicit estimator
   python scripts/trains/train.py --task=go1_pos_estimator_rough --max_iterations=5000 
   
   # BAS agile policy w/o explicit estimator (w/ implicit estimator)
   python scripts/trains/train.py --task=go1_pos_implicit_estimator_rough --max_iterations=5000 
   
   # ABS agile policy
   python scripts/trains/train.py --task=go1_pos_rough --max_iterations=5000 
   
   # PPO-Lagrangian agile policy
   python scripts/trains/train.py --task=go1_pos_rough_ppo_lagrangian --max_iterations=5000
   
   # RMA-PPO-Lagrangian agile policy
   python scripts/trains/rma_train_teacher.py --task=go1_teacher_rough_ppo_lagrangian --max_iterations=5000
   python scripts/trains/rma_train_student.py --task=go1_student_rough_ppo_lagrangian --max_iterations=5000
   # Turn from `RMA` to `Action-Distill` in one code by setting `use_action_loss=True`
   
   # recovery policy
   python scripts/trains/train.py --task=go1_rec_rough --max_iterations=1000
   

6. Play the trained policy

   python scripts/plays/play_bas.py --task=go1_pos_rough
   python scripts/plays/play_bas.py --task=go1_rec_rough

7. Use the testbed, and train/test Reach-Avoid network (Taking BAS as example):

   # For following steps in 7&8, you can change `bas` in filename to [`abs`, 'rma_teacher', 'rma_student', `implicit`] for other entries
   
   # try testbed
   python scripts/testbeds/bas_testbed.py --task=go1_pos_estimator_rough [--load_run=xxx] --num_envs=1 
   
   # train RA (be patient it will take time to converge) 
   # make sure you have at least exported one policy by play.py so the exported folder exists
   python scripts/testbeds/bas_testbed.py --task=go1_pos_estimator_rough --num_envs=1000 --headless --trainRA
   
   # test RA (only when you have trained one RA)
   python scripts/testbeds/bas_testbed.py --task=go1_pos_estimator_rough --num_envs=1 --testRA
   
   # evaluate
   python scripts/testbeds/bas_testbed.py --task=go1_pos_estimator_rough --num_envs=1000 --headless [--load_run=xxx] [--testRA]
   
   # On-policy finetuning on estimator
   python scripts/testbeds/bas_finetune_testbed.py --task=go1_pos_rough --num_envs=1000 --headless [--load_run=xxx] --testRA
   

8. Sample dataset for ray-prediction network training

python scripts/trains/camrec.py --task=go1_pos_rough --num_envs=3

• Tips 1: You can edit the shift value in Line 93 and the log_root in Line 87 to collect different dataset files in parallel (so you can merge them by simply moving the files), and manually change the obstacles in env_cfg.asset.object_files in Line 63.
• Tips 2: After collecting the data, there's a template code in train_depth_resnet.py to train the ray-prediction network, but using what you like for training CV models is highly encouraged!
• Tips 3: You may change camera configs of resolution, position, FOV, and depth range in the config file Line 151.

Hardware Deployment

System Overview

• Robot: Unitree Go1 EDU
• Perception: ZED mini Camera
• Onboard Compute: Orin NX (16GB)
• LED: PSEQT LED Lights
• Power Regulator: Pololu 12V, 15A Step-Down Voltage Regulator D24V150F12

3D Print Mounts

• Orin NX mount: STL-PCMount_v2
• ZED mini mount: STL-CameraSeat and STL-ZEDMountv1

Deployment Code Installation

• Unitree Go1 SDK
• ZED SDK
• ROS Noetic
• Pytorch on a Python 3 environment

Deployment Setup

• Low Level Control Mode for Unitree Go1: L2+A -> L2+B -> L1+L2+Start

• Network Configuration for Orin NX:

  • IP: 192.168.123.15
  • Netmask: 255.255.255.0

• Convert the .pt files of agile/recovery policy, RA value to .onnx files using src/abs_src/onnx_model_converter.py

• Modify the path of (.onnx or .pt) models in publisher_depthimg_linvel.py and depth_obstacle_depth_goal_ros.py

Deployment Scripts

1. roscore: Activate ROS Noetic Envrioment
2. cd src/bas_src/: Enter the ABS scripts file
3. python publisher_depthimg_linvel.py: Publish ray prediction results and odometry results for navigation goals
4. python led_control_ros.py: Control the two LED lights based on RA values
5. python depth_obstacle_depth_goal_ros.py: Activate the Go1 using the agile policy and the recovery policy
6. python publisher_estimator.py: Publish environment estimation results (with one additional dim for timestamp) for adaptation

Deployment Controllers

• B: Emergence stop
• Default: Go1 Running ABS based on goal command
• L2: Turn left
• R2: Turn right
• Down: Back
• Up: Stand
• A: Turn around
• X: Back to initial position

Credit

If our work does help you, please consider citing us and the following works:

@misc{zhong2025bridgingadaptivitysafetylearning,
      title={Bridging Adaptivity and Safety: Learning Agile Collision-Free Locomotion Across Varied Physics}, 
      author={Yichao Zhong and Chong Zhang and Tairan He and Guanya Shi},
      year={2025},
      eprint={2501.04276},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2501.04276}, 
}

This work builds upon Agile But Safe: Learning Collision-Free High-Speed Legged Locomotion. We used codes in Legged Gym and RSL RL, based on the paper:

• Rudin, Nikita, et al. "Learning to walk in minutes using massively parallel deep reinforcement learning." CoRL 2022.