XYZPIT/specified_bread_bag_blur_tint

R1 Lite specified bread bag — blur_tint

Single inference checkpoint at step 90000. Selected by lowest arm MAE among available evaluated checkpoints for this condition; not by mixed-unit aggregate score. Real-robot stability and grasp success have not been established. Lowest average error does not imply smallest command jumps or maximum error.

Evaluation: held-out trajectories0,1,2 (all right-hand), maximum200 frames, H32 prediction/execution, no smoothing. Native stochastic inference without explicit seed. Arm MAE 0.015291038 rad, arm MSE 0.001305312 rad²; gripper MAE 2.489982416 in0–100 units. Full left/right metrics included in open_loop_evaluation.json.

Head preprocessing: blur_tint, fixed initial non-target bounding boxes; green target highlight tracked for tint condition. Wrist videos unchanged. Instruction: Pick up the green-highlighted bread bag with the {left/right} hand.

GR00T N1.6, video[0] (Head + both wrists), state left_arm6/right_arm6/left_gripper1/right_gripper1. Action order left_gripper1/right_gripper1/left_arm6/right_arm6; real14D, model padding128, horizon32. Arms trained relative to observation state; standard Gr00tPolicy decoder returns absolute targets. Do not add current state again. Grippers absolute0–100.

Independent pretrained initialization, target100k training, batch32, LR1e-4, cosine scheduler, seed42, episode sampling1.0. This export is the selected intermediate checkpoint, not necessarily the final100k weights. Optimizer/scheduler omitted. Use compatible project's GR00T N1.6 runtime and these processor/statistics files.