ImDy: Human Inverse Dynamics from Imitated Observations (ICLR 2025)

The code release is in progress.

Environment Setup

Create a conda environment from environment.yml: conda env create -f environment.yml

Data Acquisition

1. For data access, please get in touch with [email protected].

The file structure should be

- utils
- data
 |- raw_test
   |- grf.pkl
   |- pos.pkl
   |- rot.pkl
   |- torque.pkl
   |- weight.pkl
 |- raw_train
   |- ...
 |- nimble_test
   |- figure
     |- walking
   |- walking.pkl
- osim
 |- Geometry
   |- .....
 |- Rajagopal2015_passiveCal_hipAbdMoved_noArms.osim
 |- vtp_to_ply.py

- models
 |- containing SMPL models from https://smpl.is.tue.mpg.de
 |- containing Rajagopal2015 model without arm from https://addbiomechanics.org/download_data.html
- convert.py
- adb_motion_visualize.py
- main.py
- main_freeze.py
- dataset.py
- engine.py

2. Run python convert.py to convert the raw data into a different format with per-sample pickle files including axis-angle format SMPL parameters, joints, and markers. The torques stored are acquired by summing two consecutive torques in the simulation.

Checkpoint

You could download the checkpoints here.

Train

1. Run python main.py config_path=config/IDFD_mkr.yml USE_WANDB=True Timestamp=False to pre-train the ImDy model. In IDFD_mkr.yml, you should modify the data path.

    joint_tor: true
    dpath: # your data path to imdy_train #
    cls_aug: false
......

    joint_tor: true
    dpath: # your data path to imdy_test #
    cls_aug: false

2. Run python main_freeze.py config_path=config/adb_mkr.yml USE_WANDB=True Timestamp=False to train the Addbiomechanics model.

Visualization

Run python adb_motion_visualize.py to visualize the motion from Addbiomechanics Dataset frame by frame. In line 64, you could change the angles of camera to better visualize the motion.

scene.set_camera(angles=(-pi/8,pi/2+pi/4,0),distance=2.5)