wangmingxinthu/PhaForce-Dataset

PhaForce Dataset

This repository hosts the real-robot contact-rich manipulation dataset used by PhaForce: Phase-Scheduled Visual-Force Policy Learning with Slow Planning and Fast Correction for Contact-Rich Manipulation.

Project page: https://thu-wangmx.github.io/phaforce/

Contents

The dataset is converted to FTP-1 / Open-X-Tactile-style zarr keys. Each task is stored as one zarr group:

• open_drawer.zarr: 50 trajectories, 21,653 frames
• plug_in_charger.zarr: 100 trajectories, 40,095 frames
• plug_in_usb.zarr: 60 trajectories, 21,245 frames
• wiping_board.zarr: 50 trajectories, 33,877 frames

Total: 260 trajectories, 116,870 frames.

Key Mapping

• camera_main_rgb: external RGB camera, (T, 240, 320, 3)
• left_wrist_camera_rgb: wrist RGB camera, (T, 240, 320, 3)
• left_wrist_pose: TCP pose converted from 9D pose to (x, y, z, rotvec), (T, 6)
• left_arm_joints: robot arm joints, (T, 7)
• left_hand_joints: gripper width, (T, 1)
• left_hand_joints_idx: gripper scalar index, filled with 28
• left_tactile_data_wrench: 6D TCP wrench state, (T, 6, 1)
• left_tactile_area_wrench: function-area ids, filled with [15, 16, 17, 18, 19, 20]
• left_tactile_sensor_wrench: FlexivGripperTorque
• left_tactile_type_wrench: state
• sub_task_instruction: per-step task instruction

The original empty right_wrist_img stream is intentionally not exported.

Citation

@inproceedings{wang2026phaforce,
  title={PhaForce: Phase-Scheduled Visual-Force Policy Learning with Slow Planning and Fast Correction for Contact-Rich Manipulation},
  author={Wang, Mingxin and Yue, Zhirun and Lu, Renhao and Li, Yizhe and Wang, Zihan and Pan, Guoping and Dong, Kangkang and Cheng, Jun and Cheng, Yi and Liu, Houde},
  booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2026}
}