Dataset Card for TRANSIC Data

This dataset card is accompanied with the paper titled TRANSIC: Sim-to-Real Policy Transfer by Learning from Online Correction. It includes generated simulation data and real-robot human correction data for sim-to-real transfer of robotic arm manipulation policies.

Dataset Details

Dataset Description

This dataset includes two parts, 1) simulation data used in student policy distillation and 2) real-robot data used in residual policy learning.

The first part can be found in the distillation folder. We include 5 tasks in the distillation/tasks directory. For each task, we provide 10,000 successful trajectories generated by teacher policies trained with reinforcement learning in simulation. Furthermore, we also provide matched_point_cloud_scenes.h5, a seperate collection of 59 matched point clouds in simulation and the real world. We use them to regularize the point-cloud encoder during policy training.

The second part can be found in the correction_data folder. We include real-world human correction data for 5 tasks. Each task contains different number of trajectories. Each trajectory includes observations, pre-intervention actions, and post-intervention actions for residual policy learning.

• Curated by: Yunfan Jiang
• License: MIT

Dataset Sources

• Repositories: TRANSIC, TRANSIC-Envs
• Paper: TRANSIC: Sim-to-Real Policy Transfer by Learning from Online Correction

Uses

Please see our codebase for detailed usage.

Dataset Structure

Structure for distillation/tasks/*.hdf5:

data[f"rollouts/successful/rollout_{idx}/actions"]: shape (L, 7), first 6 dimensions represent end-effector's pose change. The last dimension corresponds to the gripper action.
data[f"rollouts/successful/rollout_{idx}/eef_pos"]: shape (L + 1, 3), end-effector's positions.
data[f"rollouts/successful/rollout_{idx}/eef_quat"]: shape (L + 1, 4), end-effector's orientations in quaternion.
data[f"rollouts/successful/rollout_{idx}/franka_base"]: shape (L + 1, 7), robot base pose.
data[f"rollouts/successful/rollout_{idx}/gripper_width"]: shape (L + 1, 1), gripper's current width.
data[f"rollouts/successful/rollout_{idx}/leftfinger"]: shape (L + 1, 7), left gripper finger pose.
data[f"rollouts/successful/rollout_{idx}/q"]: shape (L + 1, 7), robot joint positions.
data[f"rollouts/successful/rollout_{idx}/rightfinger"]: shape (L + 1, 7), right gripper finger pose.
data[f"rollouts/successful/rollout_{idx}/{obj}"]: shape (L + 1, 7), pose for each object.

Structure for distillation/matched_point_cloud_scenes.h5:

# sim
data[f"{date}/{idx}/sim/ee_mask"]: shape (N,), represent if each point in the point cloud corresponds to the end-effector. 0: not end-effector, 1: end-effector.
data[f"{date}/{idx}/sim/franka_base"]: shape (7,), robot base pose.
data[f"{date}/{idx}/sim/leftfinger"]: shape (7,), left gripper finger pose.
data[f"{date}/{idx}/sim/pointcloud"]: shape (N, 3), synthetic point cloud.
data[f"{date}/{idx}/sim/q"]: shape (9,), robot joint positions, last two dimensions correspond to two gripper fingers.
data[f"{date}/{idx}/sim/rightfinger"]: shape (7,), right gripper finger pose.
data[f"{date}/{idx}/sim/{obj}"]: shape (7,), pose for each object.

# real
data[f"{date}/{idx}/real/{sample}/eef_pos"]: shape (3, 1), end-effector's position.
data[f"{date}/{idx}/real/{sample}/eef_quat"]: shape (4), end-effector's orientations in quaternion.
data[f"{date}/{idx}/real/{sample}/fk_finger_pointcloud"]: shape (N, 3), point cloud for gripper fingers obtained through forward kinematics.
data[f"{date}/{idx}/real/{sample}/gripper_width"]: shape (), gripper width.
data[f"{date}/{idx}/real/{sample}/measured_pointcloud"]: shape (N, 3), point cloud captured by depth cameras.
data[f"{date}/{idx}/real/{sample}/q"]: shape (7,), robot joint positions.

Structure for correction_data/*/*.hdf5:

data["is_human_intervention"]: shape (L,), represent human intervention (1) or not (0).
data["policy_action"]: shape (L, 8), simulation policies' actions.
data["policy_obs"]: shape (L, ...), simulation policies' observations.
data["post_intervention_eef_pose"]: shape (L, 4, 4), end-effector's pose after intervention.
data["post_intervention_q"]: shape (L, 7), robot joint positions after intervention.
data["post_intervention_gripper_q"]: shape (L, 2), gripper fingers' positions after intervention.
data["pre_intervention_eef_pose"]: shape (L, 4, 4), end-effector's pose before intervention.
data["pre_intervention_q"]: shape (L, 7), robot joint positions before intervention.
data["pre_intervention_gripper_q"]: shape (L, 2), gripper fingers' positions before intervention.

Dataset Creation

distillation/tasks/*.hdf5 are generated by teacher policies trained with reinforcement learning in simulation. distillation/matched_point_cloud_scenes.h5 and correction_data/*/*.hdf5 are manually collected in the real world.

Citation

BibTeX:

@article{jiang2024transic,
  title   = {TRANSIC: Sim-to-Real Policy Transfer by Learning from Online Correction},
  author  = {Yunfan Jiang and Chen Wang and Ruohan Zhang and Jiajun Wu and Li Fei-Fei},
  year    = {2024},
  journal = {arXiv preprint arXiv: Arxiv-2405.10315}
}

Dataset Card Contact

Yunfan Jiang, email: yunfanj[at]cs[dot]stanford[dot]edu