The official implementation of Latent Conditioned Loco-Manipulation using Motion Priors by Maciej Stępień, Rafael Kourdis, Constant Roux and Olivier Stasse (accepted for the 2025 IEEE-RAS 24th International Conference on Humanoid Robots (Humanoids)).
First install the Isaac Lab version v1.4.1 with rlgames v1.6.1. Then, in the main directory of this repository, run:
pip install -e exts/ase_envs
Training has two stages: low-level controller (LLC) - a latent conditioned motion policy that can be reused, and high-level controller (HLC) - a point reaching task that demonstrates locomotion and manipulation capabilities of the motion policy.
Below you can find an example for training Solo synthetic motions with ASE DRAIL (scripts with more examples of training and running can be found in the utils directory):
EXPERIMENT_ID=pedipulation_0
python ase/run.py --task Solo-ASE-DRAIL --motion_file ase/data/motions/solo/dataset.yaml --headless --experiment_name="solo_ase_drail" --experiment_id="$EXPERIMENT_ID"
python ase/run.py --task Solo-Pedipulation-HRL-ASE-DRAIL --motion_file ase/data/motions/solo/dataset.yaml --headless --llc_checkpoint="logs/rl_games/solo_ase_drail/solo_ase_drail_$EXPERIMENT_ID/nn/solo_ase_drail.pth" --experiment_name="solo_hrl_ase_drail" --experiment_id="$EXPERIMENT_ID"
In general, a basic template for training LLC looks like this:
python ase/run.py --task {ROBOT_NAME}-{METHOD_NAME} --motion_file {DATASET_NAME}
ROBOT_NAME:
SoloH1
METHOD_NAME:
ASEASE-CaTASE-DRAILASE-DRAIL-CaT
DATASET_NAME:
ase/data/motions/solo/dataset.yaml(Solo synthetic)ase/data/motions/solo_mocap/dataset_valid.yaml(Solo retargetted from dog motions)ase/data/motions/h1/dataset.yaml(H1 synthetic)
Template for training HLC:
python ase/run.py --task {TASK_NAME}-HRL-{METHOD_NAME} --motion_file {DATASET_NAME} --llc_checkpoint={PATH_TO_LLC_CHECKPOINT}
TASK_NAME:
Solo-PedipulationH1-Manipulation
This work was based on ASE, DRAIL, CaT and IsaacLab.
@inproceedings{mstepien2025lacoloco,
title={Latent Conditioned Loco-Manipulation using Motion Priors},
author={Maciej Stępień and Rafael Kourdis and Constant Roux and Olivier Stasse},
booktitle={IEEE-RAS 24th International Conference on Humanoid Robots (Humanoids)},
year={2025}
}