Minimal environment for quick prototyping with Cassie

This repo provides a Cassie model and a few helpful utilities to jump-start pure-python prototpying of robotics algorithms using a full scale biped example and pydrake.

Setup

A virtual environment is recommended. Any OS supported by Drake v1.28.0 is supported using the supplied requirements.txt file. If you are on a Linux distro which is not supported by the requested Drake version, you will likely have good luck simply calling pip install drake

Ubuntu 22.04 or macOS Ventura/Sonoma (arm only)

    python3 -m venv venv
    source venv/bin/activate
    pip install -r requirements.txt

Other operating systems (Untested)

    python3 -m venv venv
    source venv/bin/activate
    pip install drake

Walking data

We include trajectories of Cassie walking from a few hardware experiments. The trajectories include the floating base state (estimated using the contact-aided invariant EKF Hartley 2019), joint positions and velocities, and measured motor torques.

The data was collected using an updated version of the MPC walking controller from Bipedal Walking on Constrained Footholds with MPC Footstep Control .

If you find the data helpful, please consider citing us:

@inproceedings{Acosta2023,
  title = {Bipedal Walking on Constrained Footholds with MPC Footstep Control},
  author = {Acosta, Brian and Posa, Michael},
  year = {2023},
  month = dec,
  booktitle = {IEEE-RAS International Conference on Humanoid Robotics},
  youtube = {aTI6s2a3JSg},
  arxiv = {2309.07993},
  url = {https://ieeexplore.ieee.org/abstract/document/10375170},
  doi = {10.1109/Humanoids57100.2023.10375170}
}

The trajectories are stored in compressed .npz files and can be loaded in as shown in the example script: dynamics_examples.py

The state indices are defined below, the description of the joint dofs are defined in the file cassie_v2.urdf. The joint ordering is defined using the Drake depth-first convention when loading in the urdf. There is not an encoder on either ankle_spring_joint on the physical robot Cassie.

We estimate the joint position using inverse kinematics in our state estimator and include it in the trajectory data; however, we leave the velocity of those joints at 0.

Positions:

Index	Name
0	base_qw
1	base_qx
2	base_qy
3	base_qz
4	base_x
5	base_y
6	base_z
7	hip_roll_left
8	hip_yaw_left
9	hip_pitch_left
10	knee_left
11	knee_joint_left
12	ankle_joint_left
13	ankle_spring_joint_left
14	toe_left
15	hip_roll_right
16	hip_yaw_right
17	hip_pitch_right
18	knee_right
19	knee_joint_right
20	ankle_joint_right
21	ankle_spring_joint_right
22	toe_right

Velocities:

Index	Name
0	base_wx
1	base_wy
2	base_wz
3	base_vx
4	base_vy
5	base_vz
6	hip_roll_leftdot
7	hip_yaw_leftdot
8	hip_pitch_leftdot
9	knee_leftdot
10	knee_joint_leftdot
11	ankle_joint_leftdot
12	ankle_spring_joint_leftdot
13	toe_leftdot
14	hip_roll_rightdot
15	hip_yaw_rightdot
16	hip_pitch_rightdot
17	knee_rightdot
18	knee_joint_rightdot
19	ankle_joint_rightdot
20	ankle_spring_joint_rightdot
21	toe_rightdot

Inputs:

Index	Name
0	hip_roll_left_motor
1	hip_roll_right_motor
2	hip_yaw_left_motor
3	hip_yaw_right_motor
4	hip_pitch_left_motor
5	hip_pitch_right_motor
6	knee_left_motor
7	knee_right_motor
8	toe_left_motor
9	toe_right_motor