poolvarine/trossen-robot-data

Trossen Robot Cube Stacking Dataset

Dataset for training robotic manipulation policies on the Trossen WidowX Dual Robot cube stacking task using MuJoCo simulation.

Dataset Overview

This dataset contains demonstrations for a cube stacking task with delta actions, suitable for:

• Behavioral Cloning (BC): Dual-arm 14D action space
• Reinforcement Learning (RL): Single-arm 7D action space

---

RL Single-Arm Dataset (rl_delta_single_arm)

File: cube_stacking/rl_delta_single_arm/Trossen_windowX_DualRobot_only_right_arm_cube_stacking.pkl

Specifications

Action Space (7D) - Delta commands for RIGHT robot only:

• action[0:3]: RIGHT end-effector position delta (x, y, z) in meters
• action[3:6]: RIGHT end-effector orientation delta (angle-axis representation)
• action[6]: RIGHT gripper delta (binarized: open/close)

State Space (8D) - Absolute observations:

• state[0:3]: RIGHT end-effector absolute position (MuJoCo world frame)
• state[3:7]: RIGHT end-effector absolute orientation (quaternion)
• state[7]: RIGHT gripper absolute value [0.0, 0.044]

Image Observations:

• 4 cameras: cam_high, cam_low, cam_left_wrist, cam_right_wrist
• Resolution: 128x128 RGB
• Both robots visible in images (LEFT robot masked in action/state space)

Data Format

import pickle

with open('dataset.pkl', 'rb') as f:
    transitions = pickle.load(f)

# Each transition contains:
transition = {
    'observations': {
        'state': (8,),  # RIGHT robot state
        'cam_high': (128, 128, 3),
        'cam_low': (128, 128, 3),
        'cam_left_wrist': (128, 128, 3),
        'cam_right_wrist': (128, 128, 3),
    },
    'actions': (7,),  # RIGHT robot delta actions
    'rewards': float,
    'dones': bool,
    'masks': float,  # 1.0 - terminated
}

Download

# Install huggingface-hub
pip install huggingface-hub

# Download dataset
from huggingface_hub import hf_hub_download

file_path = hf_hub_download(
    repo_id="poolvarine/trossen-robot-data",
    filename="cube_stacking/rl_delta_single_arm/Trossen_windowX_DualRobot_only_right_arm_cube_stacking.pkl",
    repo_type="dataset"
)

# Load dataset
import pickle
with open(file_path, 'rb') as f:
    transitions = pickle.load(f)
    
print(f"Loaded {len(transitions)} transitions")

---

Usage

Training RL Policy (SERL DrQ)

# Clone SERL repository
git clone https://github.com/rail-berkeley/serl.git
cd serl/examples/async_trossen_cube_stacking/rl_delta_single_arm

# Download dataset (see above)

# Train policy
bash run_learner_trossen.sh  # Start learner
bash run_actor_trossen.sh    # Start actor (in separate terminal)

Environment Setup

from trossen_arm_mujoco.envs.cube_stacking import single_arm_rl_env

# Create environment
env = single_arm_rl_env.SERLGymWrapper1(
    env=dm_env,
    state_obs_dim=8,
    action_dim=7,
    control_mode="delta",
    action_scale=[0.025, 0.1, 0.005],  # [position, rotation, gripper]
)

---

Citation

If you use this dataset, please cite:

@misc{trossen_cube_stacking_2026,
  title={Trossen Robot Cube Stacking Dataset},
  author={Your Name},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/poolvarine/trossen-robot-data}}
}

---

License

[Specify your license here]

Contact

For questions or issues, please open an issue on the SERL repository.