license: cc-by-4.0
pretty_name: DreamTac
size_categories:
- 100K<n<1M
tags:
- robotics
- tactile-sensing
- robot-manipulation
- multimodal
- zarr
DreamTac
DreamTac is a contact-rich robot manipulation dataset collected at PKU with a Franka Emika Panda, two synchronized RGB cameras, and two Xense Photon vision-based tactile sensors mounted on the gripper fingertips.
This release follows the Open-X-Tactile submission format.
Dataset summary
- 12 tasks
- 703 successful trajectories
- 506,078 frames
- 20 FPS (approximately 7.03 hours)
- Third-person RGB:
camera_main_rgb, 224 x 224 RGB - Wrist RGB:
right_wrist_camera_rgb, 224 x 224 RGB - Tactile RGB:
right_tactile_data_gripper, two 224 x 224 RGB streams - End-effector pose: position in meters plus rotation vector in radians, expressed in the Franka base frame
- Gripper: one scalar opening value; arm joint angles are not included
The tactile area IDs are 0 for the left fingertip and 1 for the right
fingertip. Files from the source no_tactile batch are excluded.
Files
Each task is distributed as one archive under zarr_archives/ to avoid
placing roughly 100,000 individual Zarr chunk files in the Hub repository.
Extract an archive with:
tar -xf zarr_archives/pick_bread.zarr.tar
The extracted directory is a standard Zarr v2 group. Submission metadata,
hardware images, conversion manifests, and Open-X-Tactile check reports are
under submission/.
Tasks
clean_whiteboard, cut_banana, insert_screw, insert_usb, pick_bread,
pick_hongzhong, pick_place, pick_usb, push_hongzhong_pos1,
push_hongzhong_pos2, push_hongzhong_pos3, and stack_bowl.
License
DreamTac is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
References
- Paper: Dream-Tac: A Unified Tactile World Action Model for Contact-Rich Robot Manipulation
- Project repository: LYFCLOUDFAN/Dream-Tac
Contact: e1373933@u.nus.edu