mlshehab/forcesight-demos

ForceSight Demos (6_3)

Teleoperated demonstration datasets in LeRobot format (v2.1) for the ForceSight manipulation project. Recorded on a Franka Panda with a parallel-jaw hand and paired vision + tactile sensing.

Companion fine-tuned pi0.5 policies: mlshehab/forcesight.

Tasks

Four manipulation tasks, teleoperated:

• balance — balance a white box on top of blue beams.
• gear — insert a gear onto a peg.
• medicine — pick up the bottle and place it in the orange bin if full, the blue bin if empty.
• plug — insert a plug into a socket.

Subsets

All three subsets cover the same four tasks (400 episodes, 101,998 frames each). They differ in observation content / labeling:

Config	Episodes	Frames	Description
tactile_6_3	400	101,998	Full observation set incl. both tactile streams
pi05_6_3	400	101,998	pi0.5 baseline (vision + state, no tactile streams)
tapvla_6_3	400	101,998	TAP-VLA variant (vision + state, no tactile streams) —

Format

• codebase_version: v2.1
• robot_type: panda_hand
• fps: 15
• Note: image features are stored as image (not video-encoded), so parquet files are large. Load with the standard LeRobot dataset loader.

Observations

tactile_6_3 includes all five image streams below; pi05_6_3 and tapvla_6_3 include only the three camera streams (no *_touch_image).

Key	Type	Shape
left_shoulder_image	image	256×256×3
right_shoulder_image	image	256×256×3
wrist_image	image	256×256×3
left_touch_image	image	240×320×3
right_touch_image	image	240×320×3
state	float32	(8,) joint states

Tactile images captured with a GelSight sensor.

Actions

actions: float32 (8,) — joint actions.

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("mlshehab/forcesight-demos", root="tactile_6_3")
print(ds[0].keys())

License

Apache-2.0.