gr00t-n1.6-aloha-kitchen-task-2

GR00T N1.6-3B — ALOHA "Kitchen Task 2" (wipe & clear the table), 25 Hz / H=50

Single-task fine-tune on a real stationary ALOHA, sibling to gr00t-n1.6-aloha-kitchen-task-1 (the table-setting task) and to gr00t-n1.6-aloha-kitchen-task-12 (the two Kitchen tasks trained together). No task-success evaluation has been run on this checkpoint — this card makes no success claim.

• Task (verbatim from the dataset's meta/tasks.jsonl): "Wipe a plate with the sponge, then clear the table by placing both plates, the spoon, the fork, and the cup into the gray bin."
• 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 dataset Kitchen_Task_2, 51 episodes / 56,483 frames @ 25 fps, 4 cameras (high, low, wrist_left, wrist_right) at 480x640 h264, 204 videos, 1 task string, 1.5 GB. No held-out split — all 51 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

• Model-only keep copies every 10k steps; DeepSpeed 0.17.5, save_steps 5,000. Final train_loss 0.016674, train_runtime 39,693.8 s (1.512 steps/s), epoch 1.0. Inference files only are published (no optimizer / scheduler / trainer state / RNG state).
• Published checkpoints: 10k, 20k, 30k, 40k, 50k, 60k under r25_h50/checkpoint-<step>/.
• Identical recipe to the Kitchen_Task_1 repo above, differing only in the dataset and in nothing else.

Training lineage

One Slurm job, no resume. gpu26 job 15765 bo_kitchen2_r25_h50_60k ran 0 -> 60,000 steps in a single attempt (11 h 05 m on 4 GPUs) and COMPLETED; trainer_state.json of the final checkpoint reads global_step 60000. There was no restart, no mid-run recipe or loader change, and no checkpoint in this repo was produced under different settings from any other.

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.

Layout: r25_h50/checkpoint-<step>/ (+ r25_h50/aloha_config.py, the modality config the run actually trained with), modality.json (the Kitchen_Task_2 dataset's meta/modality.json), publication_manifest.json (per-file sizes + sha256 of everything published).

Provenance: gpu26 Slurm job 15765 bo_kitchen2_r25_h50_60k, COMPLETED 2026-09-16 05:50:51 UTC (11 h 05 m, 4 GPUs).