zchn/robo_test_green_grip_DG7

Green grip dataset DG7

DG7 is a LeRobot dataset (format v3.0) for one skill on an SO-101 arm: grip the green box. Each episode starts with the gripper above or near a 32 mm green box and ends after the box is lifted.

Episodes	135
Frames	46,598 at 10 fps (about 78 min)
Robot	SO-101 follower (so_follower), 6 joints
State and action	6 joint positions (shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper)
Cameras	handeye (wrist camera) and front (a camera on the leader arm), 640 x 480
Task string	grip the green box

Sources

Source	Episodes	What it is
Scripted demos	73	the grip phase of scripted box-to-marker demos
Mid-grasp demos	6	scripted grips that start near the box
Correction demos	14	scripted grips from hover positions with errors on both sides
DAgger round 1	13	a policy runs, then a scripted correction finishes the grip (only the correction is recorded)
Symmetric demos	6	scripted grips with offsets to both sides
DAgger round 2	23	as round 1, with a later policy (PG14)

The previous version, DG6 (zchn/robo_test_green_grip_DG6), is the first 112 episodes. DG7 adds the 23 DAgger round-2 episodes.

Camera views

The leader arm holds the front camera at one of two poses (V1, V2) in all episodes. A third pose (V3) is kept out of the data to test the policies on a new view. The 9 fixed test positions of the policy tests are also kept out of the data.

Models trained on it

• zchn/smolvla_green_grip_PG20: SmolVLA base, 30,000 steps; 9 of 9 test positions lifted (7 recorded + 2 after a gripper overload, see the model card).

Limits

• One arm, one table, one box, one room, mostly artificial light.
• Scripted demos: the motions are smooth and regular, not human demos.
• The DAgger episodes start where a policy stopped, so they cover a policy's errors, not the full grip.