VOFA is developed on top of the Booster Gym repository developed by Booster Robotics.
Follow these steps to set up your environment:
-
Create an environment with Python 3.8:
$ conda create --name vofa python=3.8 $ conda activate vofa
-
Install PyTorch with CUDA support:
$ conda install numpy=1.21.6 pytorch=2.0 pytorch-cuda=11.8 -c pytorch -c nvidia
-
Install Isaac Gym
Download Isaac Gym from NVIDIA’s website.
Extract and install:
$ tar -xzvf IsaacGym_Preview_4_Package.tar.gz $ cd isaacgym/python $ pip install -e .
Configure the environment to handle shared libraries, otherwise cannot found shared library of
libpython3.8:$ cd $CONDA_PREFIX $ mkdir -p ./etc/conda/activate.d $ vim ./etc/conda/activate.d/env_vars.sh # Add the following line export OLD_LD_LIBRARY_PATH=${LD_LIBRARY_PATH} export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib $ mkdir -p ./etc/conda/deactivate.d $ vim ./etc/conda/deactivate.d/env_vars.sh # Add the following line export LD_LIBRARY_PATH=${OLD_LD_LIBRARY_PATH} unset OLD_LD_LIBRARY_PATH
-
Install Python dependencies:
$ pip install -r requirements.txt
-
Download the pretrained VOFA policies from Hugging Face:
$ huggingface-cli download zichao22/vofa --repo-type model --local-dir vofa/logs/vofa
This creates the following checkpoint layout:
vofa/logs/vofa/ ├── student_policy/ │ ├── config.yaml │ └── nn/ │ └── model.pth └── teacher_policy/ ├── config.yaml └── nn/ └── model.pth
To test a trained policy in Isaac Gym, run:
$ cd vofa
$ python run.py --alg <RUNNER_NAME> --test --config <PATH/TO/CONFIG_FILE> --checkpoint <PATH/TO/CHECKPOINT>For example, to test the pretrained teacher policy:
$ python run.py --alg FalconPPO --test --config logs/vofa/teacher_policy/config.yaml --checkpoint logs/vofa/teacher_policy/nn/model.pthTo test the pretrained student policy:
$ python run.py --alg FalconCamEvaluate --test --config logs/vofa/student_policy/config.yaml --checkpoint logs/vofa/student_policy/nn/model.pthVideo recording is controlled by viewer.record_video in the config file. When enabled, videos are saved under the checkpoint directory in ../videos/<date-time>/video.mp4.
We provide a pre-trained checkpoint for the low-level whole body controller used in this work here for Booster T1 23Dof.
To train the teacher policy, run the following command:
$ cd vofa
$ python run.py --alg FalconPPO --config configs/train_teacher.yamlTo train the student policy, run the following command:
$ cd vofa
$ python run.py --alg FalconDAgger --config configs/train_student.yaml --expert_path <PATH/TO/TEACHER_CHECKPOINT>Training logs and saved models will be stored in logs/<date-time>/.
Training settings are loaded from configs/**/<task>.yaml. You can also override config values using command-line arguments:
--checkpoint: Path of the model checkpoint to load (set to-1to use the most recent model).--num_envs: Number of environments to create.--headless: Run headless without creating a viewer window.--sim_device: Device for physics simulation (e.g.,cuda:0,cpu).--rl_device: Device for the RL algorithm (e.g.,cuda:0,cpu).--seed: Random seed.--max_iterations: Maximum number of training iterations.
To add a new task, create a config file in envs/ and register the environment in envs/__init__.py.
In progress...
@misc{hu2026vofavisualobjectgoal,
title={VOFA: Visual Object Goal Pushing with Force-Adaptive Control for Humanoids},
author={Zichao Hu and Zifan Xu and Dongsik Chang and He Yin and Linh Tran and Roberto Martín-Martín and Peter Stone and Jingyu Qiao and Joydeep Biswas},
year={2026},
eprint={2605.01518},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2605.01518},
}Except for third-party material, rights held by the VOFA Authors in this repository are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).
VOFA is built on Booster Gym. Material derived from Booster Gym remains available under the Apache License 2.0, and the CC BY-NC 4.0 license does not restrict rights granted directly by that license. See NOTICE for the complete scope and third-party attributions.