gr00t-n1.6-aloha-pick-cubes-into-cup-v2

GR00T N1.6-3B — ALOHA "pick two blue cubes into the cup", v2 (50 Hz/H50 and 25 Hz/H32)

Third fine-tune campaign on the same real-robot task as gr00t-n1.6-aloha-pick-cubes-into-cup (50 Hz, H=32) and gr00t-n1.6-aloha-pick-cubes-into-cup-15hz (15 Hz, H=16).

• Task: "Pick up the two blue cubes and place them in the cup." on a stationary ALOHA (bimanual, 14-dim state/action: right_arm 0:6, right_gripper 6:7, left_arm 7:13, left_gripper 13:14).
• Data: 68 episodes / 36,390 frames recorded at 50 Hz, 4 cameras (high, low, wrist_left, wrist_right) at 240x424 h264; all 68 episodes are used for training (no held-out split in this campaign).
• Two arms, identical recipe except the data rate / action horizon:

• Arms are RELATIVE (delta w.r.t. the current joint state), grippers ABSOLUTE: pass the current 14-dim state with every call and run the checkpoint at the arm's control rate (50 Hz or 25 Hz).
• State dropout 0.3 during training (state token replaced by a learned mask token in 30 % of the samples). config.json in every checkpoint here has state_dropout_prob set to 0.0 for inference (the action_head.mask_token weight is then an ignored key); the dropout only affects the training forward pass anyway.
• Recipe: base nvidia/GR00T-N1.6-3B, embodiment tag NEW_EMBODIMENT, 4xA100-40GB, effective batch 64 (16/GPU), lr 1e-4, warmup 0.05, weight decay 1e-5, cosine, color jitter (0.3/0.4/0.5/0.08), images resized to shortest edge 256, 40k steps, checkpoints every 10k. Inference files only (no optimizer / scheduler / trainer state).
• Checkpoints are added as they land (10k, 20k, 30k, 40k per arm).

Layout: <arm>/checkpoint-<step>/ (+ <arm>/aloha_config.py), modality.json (the 50 Hz dataset's meta/modality.json, identical for the 25 Hz copy), make_25hz_dataset.py (how the 25 Hz data was derived).