happyhappy-jun/gr00t-n1.6-aloha-kitchen-task-1

GR00T N1.6-3B — ALOHA "Kitchen Task 1" (table setting), 25 Hz / H=50

Single-arm fine-tune campaign on a real stationary ALOHA, sibling to gr00t-n1.6-aloha-pick-cubes-into-cup-v2.

• Task (verbatim from the dataset's meta/tasks.jsonl): "Set the table for a meal by placing both plates on the placemat, the cup on the coaster, and the spoon and fork on the placemat beside the plates. Place the food items on the plates."
• Robot: stationary ALOHA (aloha_stationary), bimanual, 14-dim state/action: right_arm 0:6, right_gripper 6:7, left_arm 7:13, left_gripper 13:14.
• Data: LeRobot v2.1, 50 episodes / 59,639 frames @ 25 fps, 4 cameras (high, low, wrist_left, wrist_right) at 480x640 h264, 200 videos, 1 task string, 1.5 GB. No held-out split — all 50 episodes are trained on.
• Arms are RELATIVE (delta w.r.t. the current joint state), grippers ABSOLUTE: pass the current 14-dim state with every call.
• State dropout 0.3 during training. config.json in every published checkpoint has state_dropout_prob set to 0.0 for inference (the action_head.mask_token weight is then an ignored key).

Recipe

arm	base	embodiment tag	hardware	effective batch	lr / schedule	augmentation	steps
r25_h50	nvidia/GR00T-N1.6-3B (local pin GR00T-N1.6-3B-d0814e7)	NEW_EMBODIMENT	4xA100-40GB	64 (16/GPU, grad_accum 1)	1e-4, warmup 0.05, weight decay 1e-5, cosine	color jitter (0.3/0.4/0.5/0.08)	60,000

• Model-only keep copies every 10k steps; DeepSpeed 0.17.5. Final train_loss 0.01655, train_runtime 39,823 s (~1.51 steps/s). Inference files only (no optimizer / scheduler / trainer state).
• Published checkpoints: 10k, 20k, 30k, 40k, 50k, 60k under r25_h50/checkpoint-<step>/.

Deploy note

• Run the policy at 25 Hz control. Each call returns a 50-step action chunk = 2.0 s of motion.
• Arm entries are relative joint targets — add them to the current measured joint positions (use_relative_action-style); gripper entries are absolute.
• Contrast with the v2 pick-cubes repo: there r50_h50 = 50 Hz / H=50 / 1.0 s and r25_h32 = 25 Hz / H=32 / 1.28 s. This model's chunk therefore covers 2x the wall-clock horizon of the v2 R50 arm at half the control rate, so a downstream controller tuned on v2 must change its chunk duration and replan cadence.

Layout: r25_h50/checkpoint-<step>/ (+ r25_h50/aloha_config.py), modality.json (the Kitchen_Task_1 dataset's meta/modality.json), publication_manifest.json (per-file sizes + sha256 of everything published).

Provenance: gpu26 Slurm job 15210 bo_kitchen1_r25_h50_60k, COMPLETED 2026-09-15 11:43:47 KST (11 h 07 m).