SteveNguyen/sugarcube_in_mug_graspproj

grabette-sugar-cup-2008 — grasp-projected training dataset

Training-ready LeRobot v3.0 dataset for the Grabette handheld gripper, prepared from chouziel/grabette-sugar-cup-2008.

• 150 episodes, 23046 frames, 50 fps, single task: "pick up the sugar cup"
• Camera: observation.images.cam0 — 480×360, wrist view
• action (11D): Cartesian camera-local deltas dx dy dz + 6D rotation dr6d_0…5 + gripper strategy, closure
• observation.state (2D): gripper strategy, closure

Gripper representation — read this before training

The two gripper channels are projected to (strategy, closure), not raw joint angles. A policy trained on raw demonstrated angles under-closes: the recorded angle is where the human's fingers sat while pressing the object, so a position servo replaying it stops short and grips nothing.

Evaluate with evaluate.py --grasp_projection on so the policy's last two outputs are decoded back to angles before reaching the servo.

Preparation

./run_pipeline.sh chouziel/grabette-sugar-cup-2008 \
    --cameras right_cam0 --grasp-projection

Steps: reject episodes with unrecoverable SLAM loss → camera-local deltas + per-frame despike → 480×360 re-encode → grasp projection. Every episode was decode-checked after re-encoding. The source camera stream right_cam0 was renamed to cam0 to match the training and evaluation tooling.