ContactWorld

ContactWorld is a multimodal benchmark and dataset for learning predictive representations and world models for contact-rich robotic manipulation. This repository contains 12 simulated ContactWorld tasks converted to the FTP-1-compatible Open-X-Tactile (OXT) Zarr format.

• Project page: https://contact-world.github.io/
• Paper: https://arxiv.org/abs/2606.13877

Tasks

Task	Instruction
disassembly_barbed_flat	Pull the flat barbed plug out of the socket until it is fully disconnected.
disassembly_barbed_spike	Pull the spiked barbed plug out of the socket until it is fully disconnected.
disassembly_lidded_loose	Remove the loose lid from the container until the two parts are fully separated.
exploration_search	Explore the workspace through contact and locate the target object.
exploration_sorting_dim	Explore and sort the objects into their corresponding target regions under dim lighting.
exploration_sorting_normal	Explore and sort the objects into their corresponding target regions under normal lighting.
insertion_peg	Align the peg with the hole and insert it until it is fully seated.
insertion_power	Align the power plug with the socket and insert it until it is fully connected.
insertion_usb	Align the USB plug with the port and insert it until it is fully connected.
screwing_bulb	Pick up the light bulb and screw it into the socket until it is fully installed.
screwing_nut	Align the nut with the threaded bolt and screw it on until it is fully tightened.
screwing_valve	Turn the valve along its threaded axis until it reaches the target tightened state.

Robot and sensor

• Robot: Franka Panda
• End effector: parallel-jaw gripper
• Tactile sensor: GelSightWedgeR1.5
• Tactile functional-area ID: 0
• Gripper joint index: 28
• Dataset rate: 10 Hz
• Source quaternion order: [w, x, y, z]
• Exported wrist pose: [x, y, z, rx, ry, rz], using a rotation vector

Modalities

Each task Zarr contains front RGB, wrist RGB, end-effector pose, seven arm joints, one gripper joint, TacRGB, TacDepth, TacFF, per-frame language instruction, timestamps, and meta/episode_ends.

Important tactile arrays:

• right_tactile_data_tacrgb: (T, 1, 320, 240, 3), uint8, type image
• right_tactile_data_tacdepth: (T, 1, 320, 240, 1), float32, type image
• right_tactile_data_tacff: (T, 1, 10, 14, 3), float32, type matrix

TacFF is copied from the source Zarr without additional scaling or normalization. It is not asserted to be calibrated in Newtons. See supplementary_tactile_note.json.

Extraction

Each task is distributed as an independent ZIP archive:

unzip insertion_usb.zarr.zip

Then open it with Zarr:

import zarr
root = zarr.open_group("insertion_usb.zarr", mode="r")
print(root["data"]["camera_main_rgb"].shape)

Quality control

The processed dataset follows the three-step OXT-QC workflow:

1. metadata/Zarr structure precheck
2. 4D pose and trajectory visualization
3. RGB-tactile semantic review, performed manually when no VLM API is available