FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation

Learning for Dynamics & Control Conference (L4DC) 2026 Oral

[Website] [Arxiv] [Video]

TODO

Release training code
Release sim2sim code
Release sim2real code

Installation

IsaacGym Conda Env

Create mamba/conda environment, in the following we use conda for example, but you can use mamba as well.

conda create -n fcgym python=3.8
conda activate fcgym

Install IsaacGym

Download IsaacGym and extract:

wget https://developer.nvidia.com/isaac-gym-preview-4
tar -xvzf isaac-gym-preview-4

Install IsaacGym Python API:

pip install -e isaacgym/python

Test installation:

cd isaacgym/python/examples

python 1080_balls_of_solitude.py  # or
python joint_monkey.py

For libpython error:

Check conda path:
```
conda info -e
```

Set LD_LIBRARY_PATH:

export LD_LIBRARY_PATH=</path/to/conda/envs/your_env/lib>:$LD_LIBRARY_PATH

Install FALCON

git clone https://github.com/LeCAR-Lab/FALCON.git
cd FALCON

pip install -e .
pip install -e isaac_utils

Motion Retargetting

Please refer to PHC.

FALCON Training

Unitree G1_29DoF

Training Command

python humanoidverse/train_agent.py \
+exp=decoupled_locomotion_stand_height_waist_wbc_diff_force_ma_ppo_ma_env \
+simulator=isaacgym \
+domain_rand=domain_rand_rl_gym \
+rewards=dec_loco/reward_dec_loco_stand_height_ma_diff_force \
+robot=g1/g1_29dof_waist_fakehand \
+terrain=terrain_locomotion_plane \
+obs=dec_loco/g1_29dof_obs_diff_force_history_wolinvel_ma \
num_envs=4096 \
project_name=g1_29dof_falcon \
experiment_name=g1_29dof_falcon \
+opt=wandb \
obs.add_noise=True \
env.config.fix_upper_body_prob=0.3 \
robot.dof_effort_limit_scale=0.9 \
rewards.reward_initial_penalty_scale=0.1 \
rewards.reward_penalty_degree=0.0001

Evaluation Command

python humanoidverse/eval_agent.py \
+checkpoint=<path_to_your_ckpt>

After around 6k iterations, in IsaacGym:

Booster T1_29DoF

Training Command

python humanoidverse/train_agent.py \
+exp=decoupled_locomotion_stand_height_waist_wbc_diff_force_ma_ppo_ma_env \
+simulator=isaacgym \
+domain_rand=domain_rand_rl_gym \
+rewards=dec_loco/reward_dec_loco_stand_height_ma_diff_force \
+robot=t1/t1_29dof_waist_wrist \
+terrain=terrain_locomotion_plane \
+obs=dec_loco/t1_29dof_obs_diff_force_history_wolinvel_ma \
num_envs=4096 \
project_name=t1_29dof_falcon \
experiment_name=t1_29dof_falcon \
+opt=wandb \
obs.add_noise=True \
env.config.fix_upper_body_prob=0.3 \
robot.dof_effort_limit_scale=0.9 \
rewards.reward_initial_penalty_scale=0.1 \
rewards.reward_penalty_degree=0.0001 \
rewards.feet_height_target=0.08 \
rewards.feet_height_stand=0.02 \
rewards.desired_feet_max_height_for_this_air=0.08 \
rewards.desired_base_height=0.62 \
rewards.reward_scales.penalty_lower_body_action_rate=-0.5 \
rewards.reward_scales.penalty_upper_body_action_rate=-0.5 \
env.config.apply_force_pos_ratio_range=[0.5,2.0]

Evaluation Command

python humanoidverse/eval_agent.py \
+checkpoint=<path_to_your_ckpt>

After around 6k iterations, in IsaacGym:

FALCON Deploy

We provide seamless sim2sim and sim2real deployment scripts supporting both unitree_sdk2_python and booster_robotics_sdk. Please refer to this README for details.

FALCON Extension

Large Workspace

FALCON can be easily extended to larger workspace by setting larger torso command range and base height command range. We provide the sim2sim result of Unitree G1 with larger command range as an example:

https://github.com/user-attachments/assets/2d92000e-b990-45aa-a4ad-032fe0158eba

Citation

If you find our work useful, please consider citing us!

@article{zhang2025falcon,
          title={FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation},
          author={Zhang, Yuanhang and Yuan, Yifu and Gurunath, Prajwal and Gupta, Ishita and Omidshafiei, Shayegan and Agha-mohammadi, Ali-akbar and Vazquez-Chanlatte, Marcell and Pedersen, Liam and He, Tairan and Shi, Guanya},
          journal={arXiv preprint arXiv:2505.06776},
          year={2025}
        }

Other work also using FALCON's dual-agent framework:

@article{li2025softa,
          title={Hold My Beer: Learning Gentle Humanoid Locomotion and End-Effector Stabilization Control},
          author={Li, Yitang and Zhang, Yuanhang and Xiao, Wenli and Pan, Chaoyi and Weng, Haoyang and He, Guanqi and He, Tairan and Shi, Guanya},
          journal={arXiv preprint arXiv:2505.24198},
          year={2025}
        }

Acknowledgement

FALCON is built upon ASAP and HumanoidVerse.

License

This project is licensed under the MIT License - see the LICENSE file for details.