OpenViTac Benchmark Dataset

OpenViTac is a visuo-tactile robot manipulation benchmark covering physical-property perception, fragile-object handling, contact-rich manipulation, and precise insertion. This release contains 6,330 episodes in LeRobot v2.1 format: 1,500 simulated episodes across 11 tasks (1,300 with GelSight mini and 100 each with Neote and Xense for insert_USB), and 4,830 real-world episodes across 8 tasks with GelSight mini and Xense.

For the benchmark and data-collection details, see the paper and code repository.

Repository structure

Simulation/ and Real_world/ are sibling directories at the repository root. Each sensor/task directory is an independent LeRobot dataset.

OpenViTac/
├── Simulation/
│   ├── GelSight_mini/
│   ├── Neote/
│   └── Xense/
└── Real_world/
    ├── GelSight_mini/
    └── Xense/

Simulation tasks

Each Simulation/<sensor>/<task>/ directory is a separate LeRobot dataset. Its meta/ directory contains episode metadata, data/chunk-000/ contains one Parquet file per episode, and videos/chunk-000/ contains the corresponding camera and tactile videos. GelSight mini data is under Simulation/GelSight_mini/.

Capability	Task directory	Task summary	Episodes
Physical-property perception	weight_classify	Identify whether a block is light or heavy through interaction and place it on the matching target.	100
Physical-property perception	hardness_classify	Identify whether a block is soft or hard through contact and place it on the matching target.	100
Physical-property perception	roughness_classify	Identify whether a block is smooth or rough through contact and place it on the matching target.	100
Physical-property perception	roughness_guided_regrasp	Use surface roughness to decide whether to keep the grasped block or regrasp another one.	100
Physical-property perception	empty_can_selection	Distinguish an empty can from visually similar cans through interaction and place it in a basket.	100
Fragility-aware manipulation	grasp_chip	Grasp and move a fragile chip without damaging it.	100
Contact-rich manipulation	gear_assembly	Mount and rotate a gear while maintaining engagement with a paired gear.	100
Contact-rich manipulation	pull_drawer	Maintain a stable grasp on the handle while opening a drawer.	100
Precision manipulation	insert_USB	Align a USB connector with its slot and insert it.	100
Precision manipulation	insert_block_v1	Align and insert a designated block into the matching opening.	100
Precision manipulation	insert_block_v2	Select the instructed object, then insert it into the matching opening.	300

The insert_USB task also includes Neote and Xense tactile sensor data, with 100 episodes each in Simulation/Neote/insert_USB/ and Simulation/Xense/insert_USB/.

Real-world tasks

The real-world release contains 4,830 episodes, 1,723,607 frames, and 24,247 files, recorded with a Franka robot at 30 FPS. The 13 independent datasets are stored under Real_world/<sensor>/<task>/, with sensor names GelSight_mini and Xense. Episode counts below come from each dataset's meta/info.json.

Task directory	Task summary	GelSight mini episodes	Xense episodes
empty_can_selection	Find the empty can by touch and place it in the basket.	99	100
roughness_classify	Classify surface roughness by touch and place the block on the matching plate.	—	40
grasp_chip	Gently grasp a chip and place it on a plate.	—	99
gear_assembly	Assemble the red gear onto the base and rotate the gear pair.	200	400
pull_drawer	Grasp the upper drawer handle and pull the drawer open.	—	100
insert_USB	Move the USB plug from the blue slot into the red USB slot.	299	300
insert_block_v1	Insert the blue cube into the matching hole in the yellow box.	400	399
insert_block_v2	Insert the instructed blue cube, blue half cylinder, or red hexagonal prism into the matching hole.	1195	1199
Total		2,193	2,637

Real-world dataset layout

Real_world/<sensor>/<task>/
├── README.md
├── meta/
│   ├── info.json
│   ├── tasks.jsonl
│   ├── episodes.jsonl
│   ├── dataset_metadata.json
│   └── conversion_manifest.jsonl
├── data/chunk-XXX/episode_XXXXXX.parquet
└── videos/chunk-XXX/
    ├── observation.images.head/episode_XXXXXX.mp4
    ├── observation.images.wrist/episode_XXXXXX.mp4
    ├── observation.images.tactile_left/episode_XXXXXX.mp4
    └── observation.images.tactile_right/episode_XXXXXX.mp4

Chunk and episode indices are zero-padded. Some datasets span multiple chunks. The meta/ directory also contains statistics and, where applicable, exclusion manifests. Refer to the README and metadata inside each sensor/task directory for the exact task instructions, source mapping, and excluded episodes.

Observations and actions

• Four synchronized MP4 streams per episode: head and wrist RGB at 640 × 480, and left/right tactile images at 224 × 224, all at 30 FPS.
• observation.state: 8 values, comprising 7 measured Franka joint positions and the measured gripper opening state.
• Additional observations include joint positions, end-effector poses as quaternion (xyzw) and absolute rotation vector, a flattened rotation matrix, and the measured gripper state.
• action: 7 values copied from the recorded source commands, [dx, dy, dz, drot_x, drot_y, drot_z, gripper_target_position]. The final value is the commanded absolute gripper target, while observation.gripper is the measured state. Consult each dataset's meta/dataset_metadata.json for exact coordinate frames, units, and gripper semantics.

Xense tactile frames were resized from the SDK's 400 × 700 Rectify output to 224 × 224 during collection, without preserving the original aspect ratio. Native Rectify frames are not included, and the LeRobot conversion does not apply a second resize. GelSight mini tactile frames are also stored at 224 × 224.