robot-learning-vislab/fr3-tomato-in-pan

30 demonstrations of picking up a toy tomato and placing it in a pan.

Teleoperated with a GELLO leader arm through robot-stack on a Franka FR3 under a cartesian impedance controller, recorded from a ZED 2i left camera.

episodes	30
frames	3939
rate	10 Hz
robot	franka_crisp
camera	1280x720, uncropped
task string	put the tomato in the pan

Each episode starts from a randomised pose (±2 cm in xy, ±15° in yaw about a measured start pose) that the arm drives to by itself, so the starts are drawn from the same distribution a policy would be reset to. Between episodes the operator puts the object back on its mark by hand while the arm holds still.

State is 14-D (fr3_joint*.pos, ee.x..ee.rz as a base-frame position and rotation vector, gripper.pos normalised by an 80 mm opening). Actions are 7-D: six pose deltas plus a gripper command, which uses the leader trigger's own hysteresis (shuts below 0.45, opens above 0.55).

Notes

• Uncropped: the whole 1280x720 ZED frame is kept so a crop can be chosen per downstream task. The pan sits near the right edge of the frame, so check any crop still contains it.
• Each episode has two gripper transitions: a grasp, then a release over the pan. Episodes end with the tomato in the pan and the arm withdrawn.
• Episodes are long: 64-78 steps once subsampled at stride 4, against 27-53 for the pick-only sets. A policy trained with a 50-step horizon could not reach the placement.