FurnitureBench

FurnitureBench is a benchmark designed for long-horizon, complex robotic manipulation tasks using real-world furniture assembly. It provides a reproducible platform featuring:

• Long-horizon complex manipulation tasks
• Standardized environment setup
• Python-based robot control stack
• FurnitureSim: A simulated environment for testing
• Large-scale teleoperation dataset: Over 200 hours of data available

For detailed implementation guides and tutorials, refer to the official documentation.

The rolf/ package is organized as follows:

• main.py: Entry point that sets up experiments and triggers training via trainer.py.
• trainer.py: Contains the core training and evaluation logic.
• algorithms/: Implementations of RL (PPO, DDPG, TD3, SAC, Dreamer) and IL (BC, GAIL) algorithms.
• config/: Hydra-based YAML configurations.
  • algo/: Algorithm-specific hyperparameters.
  • env/: Environment-specific hyperparameters.
• networks/: Implementations of neural networks (e.g., policy and value functions).
• utils/: Helper functions.

FurnitureSim is a simulator built on top of Isaac Gym.

System Requirements:

• OS: Ubuntu
• Python: 3.8 or higher

To use FurnitureSim, you must follow the specific installation and usage guides provided in the project documentation:

1. Installation: Follow the Installing FurnitureSim guide.
2. Usage: Follow the How to Use FurnitureSim tutorial.

VIP is integrated into TorchRL as a visual representation transform for Gym environments. After installing TorchRL, you can wrap an environment with VIPTransform to extract VIP representations from pixel observations.

from torchrl.envs.transforms import VIPTransform
# Assuming my_env is a TransformedEnv or similar
env = TransformedEnv(my_env, VIPTransform(keys_in=["next_pixels"], download=True))

Use train_offline.py to train models on various environments. You must specify the --env_name and a corresponding --config file.

Locomotion (e.g., halfcheetah):

python train_offline.py --env_name=halfcheetah-medium-expert-v2 --config=configs/mujoco_config.py

AntMaze:

python train_offline.py --env_name=antmaze-large-play-v0 --config=configs/antmaze_config.py --eval_episodes=100 --eval_interval=100000

Kitchen and Adroit (e.g., pen-human):

python train_offline.py --env_name=pen-human-v0 --config=configs/kitchen_config.py

# Example for AntMaze
python train_offline.py --env_name=antmaze-large-play-v0 --config=configs/antmaze_config.py --eval_episodes=100 --eval_interval=100000

For large-scale pre-training equivalent to the released VIP model, use the config_vip_ego4d.yaml configuration file.

python train_vip.py --config-name=config_vip_ego4d dataset=ego4d datapath=ego4d_dataset_path

To set up the environment for Implicit Q-Learning, upgrade pip, install the required packages from requirements.txt, and install a JAX wheel compatible with CUDA 11 and cuDNN 8.

For other CUDA configurations, refer to the official JAX installation guide.

pip install --upgrade pip

pip install -r requirements.txt

# Installs the wheel compatible with Cuda 11 and cudnn 8.
pip install --upgrade "jax[cuda]>=0.2.27" -f https://storage.googleapis.com/jax-releases/jax_releases.html

To add a new algorithm to the framework, you must implement the following components:

• run.py: Experiment setup.
• your_config.yaml: Hyperparameter configuration (compatible with Hydra).
• your_trainer.py: Training and evaluation loops.
• your_agent.py: The algorithm implementation.
• your_rollout.py: Rollout logic.
• your_network.py: Model architectures (policy, value functions, etc.).

For a reference implementation, see the skill-chaining repository.

You can install R3M into an existing conda environment by running the editable install command from the repository root. Alternatively, you can create a fresh environment using the provided r3m_base.yaml file and then perform the editable installation.

To verify the installation, attempt to import the package in a Python shell.

# Install from an existing conda environment
pip install -e .

# Verify installation
python -c "import r3m; print('Success')"

You can train VIP from scratch using a custom video dataset. Your dataset must follow this directory structure:

my_dataset_path/
    video0/
        0.png
        1.png
        ...
    video1/
    video2/
    ...

Run the training script using train_vip.py with the appropriate dataset and datapath arguments.

python train_vip.py --config-name=config_vip dataset=my_dataset_name datapath=my_dataset_path