https://github.com/foruck/imdy
ImDy: Human Inverse Dynamics from Imitated Observations (ICLR 2025)
https://github.com/foruck/imdy
Last synced: 4 months ago
JSON representation
ImDy: Human Inverse Dynamics from Imitated Observations (ICLR 2025)
- Host: GitHub
- URL: https://github.com/foruck/imdy
- Owner: Foruck
- Created: 2024-10-23T05:44:57.000Z (over 1 year ago)
- Default Branch: master
- Last Pushed: 2025-07-04T06:10:15.000Z (11 months ago)
- Last Synced: 2025-10-06T16:50:01.314Z (8 months ago)
- Language: Python
- Homepage:
- Size: 34.9 MB
- Stars: 16
- Watchers: 2
- Forks: 2
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# [ImDy: Human Inverse Dynamics from Imitated Observations](https://foruck.github.io/ImDy/) (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 visit [our Hugging Face repository](https://huggingface.co/datasets/XinpengLiu/ImDy).
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; python generate_cand.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](https://drive.google.com/drive/folders/1kDr_UpdpE19efO99sp-oCInreX7o1CqY?usp=sharing).
## 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)
```
