tbai

Towards better athletic intelligence

Installing tbai

Install Catkin tools using the steps provided here

# Install dependencies
sudo apt install libmpfr-dev

# Download project
mkdir -p <your-workspace>/src && cd <your-file> && catkin init && cd src
git clone git@github.com:asblab-research/tbai.git --recursive

# Install other dependencies using rosdep
cd .. && rosdep install --from-paths src --ignore-src -r -y && cd src/tbai

# !! Now install libtorch by following the installation guideline below

Installing Additional Dependencies

1. Installing libtorch C++

There are two steps to installing libtorch. First, you need to download a suitable libtorch version. Once the library is downloaded, it's necessary to create a symlink to it in the dependencies folder. Here's how to do it:

1.1 Getting libtorch download link

Get your download link from the official PyTorch website. Note that opting for the (cxx11 ABI) version is paramount. If you download the (Pre-cxx11 ABI) version, things won't work as necessary.

1.2 Downloading libtorch and creating a symlink

Now that you have your url, you can download the library, unzip it and create a symlink in the dependencies folder.

wget <your-url> # Latest URL (JUL-29-2024) - https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-2.4.0%2Bcpu.zip
unzip <downloaded-zip> -d <your-workspace>/src/tbai/dependencies  # can be `dependencies` folder of the package
ln -s <your-folder>/libtorch dependencies  # Only necessary if, in the previous step, you did not unzip in `dependencies`

Your dependencies folder should not look as follows:

Installing aws-robomaker-hospital-world

cd <your-workspace>/src
git clone git@github.com:asblab-research/aws-robomaker-hospital-world.git

Virtual Enivronment / Anaconda

In case you are using a venv these additional packages will be needed.

pip install empy==3.3.4
pip install catkin_pkg
sudo apt-get install doxygen
pip install rospkg
pip install defusedxml

Building the package

#Build aws-robomaker-hospital-world
cd <your-workspace> && catkin build aws_robomaker_hospital_world

# Build tbai
cd src/tbai
catkin config -DCMAKE_BUILD_TYPE=Release
bash ./tbai.bash --build  # This will only build the necessary packages

# Source
cd ../.. && source devel/setup.bash

Usage

Currently only tbai_mpc_perceptive and tbai_rl_perceptive are modified to use aws-robomaker-hospital-world.

tbai_mpc_perceptive

# Start ROS and relevant nodes
roslaunch tbai_mpc_perceptive simple_hospital_two_floor.launch # single floor world can be launched using simple_hospital.launch

# Change controllers (in a different terminal)
rostopic pub /anymal_d/change_controller std_msgs/String "data: 'STAND'"
rostopic pub /anymal_d/change_controller std_msgs/String "data: 'WBC'"
#rostopic pub /anymal_d/change_controller std_msgs/String "data: 'SIT'"

# In the MPC controller terminal (started with roslaunch) change gait to trot

tbai_rl_perceptive

# Start ROS and relevant nodes
roslaunch tbai_rl_perceptive simple_hospital_two_floor.launch # single floor world can be launched using simple_hospital.launch

# Change controllers (in a different terminal)
rostopic pub /anymal_d/change_controller std_msgs/String "data: 'RL'"
rostopic pub /anymal_d/change_controller std_msgs/String "data: 'BOB'"  # RL and BOB are the same controllers
rostopic pub /anymal_d/change_controller std_msgs/String "data: 'STAND'"
#rostopic pub /anymal_d/change_controller std_msgs/String "data: 'SIT'"

Controlling the quadruped

A joystick is used to control the robot. joy_node is started through the launch file.

left analog stick = translation ; right analog stick = rotation

If you do not have access to a controller (recommended), a joystick emulator is provided.

rosrun tbai_rl_perceptive joy_emulator.py

Implemented controllers

📦tbai
 ┣ 📂tbai_static               # Static (high gain PD) controller
 ┣ 📂tbai_mpc_perceptive       # Perceptive NMPC controller [1]
 ┣ 📂tbai_mpc_blind            # Blind NMPC controller [1]
 ┣ 📂tbai_rl_perceptive        # Perceptive RL controller [2]
 ┣ 📂tbai_rl_blind             # Blind RL controller [2]
 ┣ 📂tbai_dtc                  # DTC controller (perceptive) [3]
 ┣ 📂tbai_joe                  # Perceptive NMPC controller with NN-based tracking controller [1],[3]

 [1] Perceptive Locomotion through Nonlinear Model Predictive Control
     https://arxiv.org/abs/2208.08373
 [2] Learning robust perceptive locomotion for quadrupedal robots in the wild
     https://arxiv.org/abs/2201.08117
 [3] DTC: Deep Tracking Control
     https://arxiv.org/abs/2309.15462

Perceptive MPC

mpc_perceptive_f.mp4

Blind MPC

mpc_blind_f.mp4

Perceptive RL

rl_perceptive_fe.mp4

Blind RL

rl_blind_fe.mp4

Dtc

dtc_f.mp4

Joe

joe_f.mp4

System architecture

Controller architectures

Mpc

Rl (Bob)

Dtc

Joe

If any of the steps throws an error for you, please let use know and we will try to extend this guideline with a fix as soon as possible. Thanks 🤗

Credits

This project stands on the shoulders of giants. None of this would have been possible were it not for many amazing open-source projects. Here are a couple that most inspiration was drawn from and that were instrumental during the development:

• https://github.com/leggedrobotics/ocs2
• https://github.com/qiayuanl/legged_control
• https://github.com/leggedrobotics/legged_gym
• https://github.com/leggedrobotics/rsl_rl
• https://github.com/ANYbotics/elevation_mapping
• https://github.com/leggedrobotics/elevation_mapping_cupy
• https://github.com/bernhardpg/quadruped_locomotion
• https://github.com/stack-of-tasks/pinocchio
• https://x-io.co.uk/open-source-imu-and-ahrs-algorithms/
• https://github.com/mayataka/robotoc
• https://github.com/mayataka/legged_state_estimator
• hundreds of others ...

Thank you all 🤗