Downloads · 30 days
0
happyhappy-jun/gr00t-n1.6-aloha-cosmos-mt-80k
gr00t-n1.6-aloha-cosmos-mt-80k is a robotics model from happyhappy-jun. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Multi-task fine-tune of GR00T N1.6-3B on a 19-task stationary-ALOHA corpus, sibling to gr00t-n1.6-aloha-kitchen-task-1 and gr00t-n1.6-aloha-pick-cubes-into-cup-v2.
Downloads · 30 days
0
Access
Public
Updated Sep 16, 2026
Repo size
52.6 GB
Likes
0
Public
Click a slice to open those files.
.safetensors52.6 GB · 100%
From the Hugging Face model README
Multi-task fine-tune of GR00T N1.6-3B on a 19-task stationary-ALOHA corpus, sibling to gr00t-n1.6-aloha-kitchen-task-1 and gr00t-n1.6-aloha-pick-cubes-into-cup-v2. This is a pretraining checkpoint: it is meant to be fine-tuned on a downstream ALOHA task, not deployed as-is. No task-success evaluation has been run on it — this card makes no success claim.
aloha_stationary), bimanual, 14-dim state/action:
right_arm 0:6, right_gripper 6:7, left_arm 7:13, left_gripper 13:14.aloha_static_cosmos_25hz_v1, 976 episodes / 373,225 frames @ 25 fps,
19 task strings, 4 cameras (high, low, wrist_left, wrist_right) at 480x640 h264,
3,904 videos. No held-out split — all 976 episodes are trained on.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).The 19 task strings the model was conditioned on (verbatim from the dataset's meta/tasks.jsonl):
| # | task |
|---|---|
| 0 | Place the battery into the slot of the remote controller. |
| 1 | Pick up the candy and unwrap it. |
| 2 | Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the 'Hot Water' and 'Travel Mug' buttons. |
| 3 | Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray. |
| 4 | Pick up the plastic cup and open its lid. |
| 5 | "Pick up the fork and place it on the plate." |
| 6 | Transfer one of the two balls in the right glass into the left glass, then transfer it back to the right glass. |
| 7 | Pick up the pencil with the right arm, hand it over to the left arm then place it back onto the table. |
| 8 | Pick up the screwdriver with the right arm, hand it over to the left arm then place it into the cup. |
| 9 | "Cut a small piece of tape from the tape dispenser then place it on the cardboard box's edge." |
| 10 | Pick up a piece of paper towel and place it on the spilled liquid. |
| 11 | Pick up the platic cup with the right arm, then pop its lid open with the left arm. |
| 12 | Pick up the platic cup with the left arm, then pop its lid open with the right arm. |
| 13 | Slide open the ziploc bag. |
| 14 | fold shirt |
| 15 | put brown chicken wing on plate |
| 16 | put candies in bowl |
| 17 | put candy in bag |
| 18 | put purple eggplant on plate |
(Task strings 5, 9 and 11-12 carry the dataset's own quoting and spelling, including the "platic" typo; they are reproduced unchanged because that is the literal conditioning text.)
| 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 | 4xH200 | 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) | 80,000 |
save_steps 5,000, model-only keep/ copies every 10k steps. Final
loss at step 80,000 (last logged training-loss sample) 0.0032; epoch 0.75.
Inference files only are published (no optimizer / scheduler / trainer state / RNG state).r25_h50/checkpoint-<step>/.The 80,000 steps were produced by four Slurm jobs on the same output directory, not one:
| attempt | steps covered | outcome |
|---|---|---|
| 1 | 0 -> 15,000 | failed at the step-15,000 save |
| 2 | resumed from 10,000, reached ~21,000 | failed with an NCCL watchdog abort |
| 3 | resumed from 20,000, reached 20,758 | cancelled deliberately |
| 4 | resumed from 20,000 -> 80,000 | COMPLETED (8 h 38 m, 4 GPUs) |
Attempts 1-3 ran with the data loader's episode_sampling_rate at 0.1; attempt 4 ran with
0.5. The change was a throughput fix, not a recipe preference: at 0.1 each episode is spread
over ~10 shards and every shard decodes all frames of all 4 cameras of every episode it touches,
so shard construction (800-1640 s) could not keep up with consumption (~450 s) and the run stalled
until the NCCL watchdog aborted it. At 0.5 the same shard covers ~6 episodes instead of ~31,
roughly 5x less decoding, and throughput went from 25-29 steps/min to a steady 120 steps/min.
The practical consequence is that steps 0-20,000 and steps 20,000-80,000 saw different episode
mixing within a batch — at 0.5 a 64-sample batch draws from fewer distinct episodes (<=24) than
at 0.1. Nothing else in the recipe changed across the four attempts: same base checkpoint, same
modality config, same dataset, same optimizer settings, same state_dropout_prob 0.3, same code
commit. checkpoint-10000 and checkpoint-20000 are therefore the only published checkpoints
trained entirely under the 0.1 regime.
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 aloha_static_cosmos_25hz_v1 dataset's
meta/modality.json), publication_manifest.json (per-file sizes + sha256 of everything published).
Provenance: gpu26 Slurm job 15852 bo_cosmos_mt_r25_h50_80k_resume_3, COMPLETED
2026-09-16 11:36:39 KST (08 h 38 m, 4 GPUs); earlier attempts 15461 / 15462 / 15703 as tabulated above.