sunuk000/act-rb5-lunchbox-packing
Robotics • 51.7M • Updated • 27
140 teleoperated demonstrations of a lunchbox-packing task on a Rainbow Robotics RB5-850, recorded with a kinematically different leader arm (RB-Y1, seven joints) and three RGB cameras. In LeRobot format (v3.0), 30 Hz.
Task: put the carrot (a model) in the black box, then the Oreo in the large hole and the RITZ in the small hole of the meal box, always in that order.
Code, videos, and the policy trained on this dataset: https://github.com/sunuk00/rb5-lunchbox-packing
| Episodes | Content |
|---|---|
| 0–48 | Carrot and one snack (Oreo or RITZ), normal demonstrations |
| 49–64 | Carrot and one snack, recovery demonstrations (one deliberate missed grasp, then a re-grasp) |
| 65–115 | All three items, normal demonstrations |
| 116–139 | All three items, recovery demonstrations (each item missed 8 times, in four directions) |
228,583 frames in total. In the episodes with one snack, the other snack is absent; the same rule applies (the task string says "the snack").
| Key | Shape | Meaning |
|---|---|---|
observation.state |
7 | RB5 joint angles J1–J6 [deg] (encoders), gripper opening (0 closed, 1 open) |
action |
4 | Commanded TCP position x, y, z [mm] in the robot base frame, gripper command (0 closed, 1 open) |
observation.images.front |
240×320×3 | Fixed camera, oblique view from the front (RealSense D435, RGB only) |
observation.images.top |
240×320×3 | Fixed camera above the item area, looking down (Logitech C270) |
observation.images.grip |
240×320×3 | Camera on the gripper, sees the fingers (Logitech Brio 101) |
CC BY 4.0