Home

Awesome

HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation

Paper Website

We present HumanoidBench, a simulated humanoid robot benchmark consisting of $15$ whole-body manipulation and $12$ locomotion tasks. This repo contains the code for environments and training.

image

Directories

Structure of the repository:

Installation

Create a clean conda environment:

conda create -n humanoidbench python=3.11
conda activate humanoidbench

Then, install the required packages:

# Install HumanoidBench
pip install -e .

# jax GPU version
pip install "jax[cuda12]==0.4.28"
# Or, jax CPU version
pip install "jax[cpu]==0.4.28"

# Install jaxrl
pip install -r requirements_jaxrl.txt

# Install dreamer
pip install -r requirements_dreamer.txt

# Install td-mpc2
pip install -r requirements_tdmpc.txt

Environments

Main Benchmark Tasks

Test Environments with Random Actions

python -m humanoid_bench.test_env --env h1hand-walk-v0

Test Environments with Hierarchical Policy and Random Actions

# Define checkpoints to pre-trained low-level policy and obs normalization
export POLICY_PATH="data/reach_two_hands/torch_model.pt"
export MEAN_PATH="data/reach_two_hands/mean.npy"
export VAR_PATH="data/reach_two_hands/var.npy"

# Test the environment
python -m humanoid_bench.test_env --env h1hand-push-v0 --policy_path ${POLICY_PATH} --mean_path ${MEAN_PATH} --var_path ${VAR_PATH} --policy_type "reach_double_relative"

Test Low-Level Reaching Policy (trained with MJX, testing on classical MuJoCo)

# One-hand reaching
python -m humanoid_bench.mjx.mjx_test --with_full_model 

# Two-hand reaching
python -m humanoid_bench.mjx.mjx_test --with_full_model --task=reach_two_hands --folder=./data/reach_two_hands

Change Observations

As a default, the environment returns a privileged state of the environment (e.g., robot state + environment state). To get proprio, visual, and tactile sensing, set obs_wrapper=True and accordingly select the required sensors, e.g. sensors="proprio,image,tactile". When using tactile sensing, make sure to use h1touch in place of h1hand. Full test instruction:

python -m humanoid_bench.test_env --env h1touch-stand-v0 --obs_wrapper True --sensors "proprio,image,tactile"

Other Environments

In addition to the main benchmark tasks listed above, you can run the following environements that feature the robot without hands:

The robot with low-dimensional hands:

And the Unitree G1 robot with three-finger hands:

Training

# Define TASK
export TASK="h1hand-sit_simple-v0"

# Train TD-MPC2
python -m tdmpc2.train disable_wandb=False wandb_entity=[WANDB_ENTITY] exp_name=tdmpc task=humanoid_${TASK} seed=0

# Train DreamerV3
python -m embodied.agents.dreamerv3.train --configs humanoid_benchmark --run.wandb True --run.wandb_entity [WANDB_ENTITY] --method dreamer --logdir logs --task humanoid_${TASK} --seed 0

# Train SAC
python ./jaxrl_m/examples/mujoco/run_mujoco_sac.py --env_name ${TASK} --wandb_entity [WANDB_ENTITY] --seed 0

# Train PPO (not using MJX)
python ./ppo/run_sb3_ppo.py --env_name ${TASK} --wandb_entity [WANDB_ENTITY] --seed 0

Training Hierarchical Policies

# Define TASK
export TASK="h1hand-push-v0"

# Define checkpoints to pre-trained low-level policy and obs normalization
export POLICY_PATH="data/reach_one_hand/torch_model.pt"
export MEAN_PATH="data/reach_one_hand/mean.npy"
export VAR_PATH="data/reach_one_hand/var.npy"

# Train TD-MPC2 with pre-trained low-level policy
python -m tdmpc2.train disable_wandb=False wandb_entity=[WANDB_ENTITY] exp_name=tdmpc task=humanoid_${TASK} seed=0 policy_path=${POLICY_PATH} mean_path=${MEAN_PATH} var_path=${VAR_PATH} policy_type="reach_single"

# Train DreamerV3 with pre-trained low-level policy
python -m embodied.agents.dreamerv3.train --configs humanoid_benchmark --run.wandb True --run.wandb_entity [WANDB_ENTITY] --method dreamer_${TASK}_hierarchical --logdir logs --env.humanoid.policy_path ${POLICY_PATH} --env.humanoid.mean_path ${MEAN_PATH} --env.humanoid.var_path ${VAR_PATH} --env.humanoid.policy_type="reach_single" --task humanoid_${TASK} --seed 0

Paper Training Curves

Please find here json files including all the training curves, so that comparing with our baselines will not necessarily require re-running them in the future.

The json files follow this key structure: task -> method -> seed_X -> (million_steps or return). As an example to access the return sequence for one seed of the SAC run for the walk task, you can query the json data as data['walk']['SAC']['seed_0']['return'].

Citation

If you find HumanoidBench useful for your research, please cite this work:

@article{sferrazza2024humanoidbench,
    title={HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation},
    author={Carmelo Sferrazza and Dun-Ming Huang and Xingyu Lin and Youngwoon Lee and Pieter Abbeel},
    journal={arXiv Preprint arxiv:2403.10506},
    year={2024}
}

References

This codebase contains some files adapted from other sources: