Downloads · 30 days
8
22% of all-time downloads
happyhappy-jun/gr00t-n1.6-ours-racer-ft-at60k
gr00t-n1.6-ours-racer-ft-at60k 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.
This is an independently trained target-task fine-tuning (FT) checkpoint of NVIDIA GR00T N1.6, fine-tuned from the Ours-RACER mid-training checkpoint happyhappy-jun/gr00t-n1.6-ours-racer-mt80k on the 18 atomic-seen (A…
Downloads · 30 days
8
22% of all-time downloads
All-time downloads
37
Public
Parameters
3.3B
6.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors6.6 GB · 100%
From the Hugging Face model README
This is an independently trained target-task fine-tuning (FT) checkpoint of
NVIDIA GR00T N1.6, fine-tuned from the
Ours-RACER mid-training checkpoint
happyhappy-jun/gr00t-n1.6-ours-racer-mt80k
on the 18 atomic-seen (AT) RoboCasa365 target tasks with full-task language only. It is
not an official NVIDIA release.
This repository is released anonymously for ICLR 2027 review.
It is the Ours-RACER member of a three-arm family that shares the target task population, the full-task-only language and the FT recipe, and differs in the mid-training parent (Ours-RACER segmentation instead of UVD-RACER or RDD-RACER). The per-arm staged copies of the target datasets differ only in a subtask-annotation overlay column that this full-task stage never reads. This checkpoint is a flat, full-task policy: it takes the whole task instruction directly and needs no high-level planner.
ft_at of the Ours arm's 20260914_target target fine-tuning harnesscheckpoint-60000 (60,000 optimizer updates)14331 on gpu26 (the lab's AWS cluster, us-west-2), a100 availability zone,
partition a100, QoS normal8, 4 GPUs on a single node. The run reached
global_step 60000 of max_steps 60000 (trainer_state.json, epoch 1.0,
best_model_checkpoint null), and its checkpoint-60000 export was probed file by file at
the source — sizes, Safetensors header-derived sizes, JSON parse, trainer_state step and
the resolved language key — before it was staged for publicationcheckpoint-80000 of run
20260913_mix50_80k/mt (job 47), published as
happyhappy-jun/gr00t-n1.6-ours-racer-mt80k,
loaded weights-only through --base-model-path with a fresh optimizer state — no
optimizer, scheduler or RNG state carried over. The base was read from an archived copy of that same MT
checkpoint-80000 (run 20260913_mix50_80k/mt) held in the lab's object store, and the
launcher waited on that run's own COMPLETE.json marker before starting.d0814e7, loaded fresh by the MT parent
(which mid-trained on 300 RoboCasa365 human-teleoperation datasets, 32,043 episodes /
29,106,226 frames, with a 50:50 full-task/subtask language mixture over the Ours
subtask overlay)f9234bf979b43fd83db77e79a816d0b8c1ccb9b462ad6f7e5ba201c188223750The model files are copied byte-for-byte from that checkpoint. No weights,
normalization statistics, modality settings, or configuration values were changed
for this upload. See publication_manifest.json for sizes and SHA-256 checksums
of the published source files.
| Item | Value |
|---|---|
| Data | the 18 official RoboCasa365 atomic-seen target LeRobot datasets (one task/date directory each), equal mix ratios (1 each); 9,126 episodes / 2,231,347 frames, counted by the staging job that materialized the private training copy |
| Tasks | CloseBlenderLid, CloseFridge, CloseToasterOvenDoor, CoffeeSetupMug, NavigateKitchen, OpenCabinet, OpenDrawer, OpenStandMixerHead, PickPlaceCounterToCabinet, PickPlaceCounterToStove, PickPlaceDrawerToCounter, PickPlaceSinkToCounter, PickPlaceToasterToCounter, SlideDishwasherRack, TurnOffStove, TurnOnElectricKettle, TurnOnMicrowave, TurnOnSinkFaucet |
| Language conditioning | full-task instructions only — the modality config (robocasa365_config.py) exposes the single language key annotation.human.task_description; no subtask language is consumed at this stage, so no segment annotation of any arm enters this run |
| Embodiment tag | NEW_EMBODIMENT |
| Steps | 60,000 (trainer_state.json: global_step 60000, max_steps 60000, epoch 1.0) |
| Effective batch | 128 (global batch 128 × gradient accumulation 1; microbatch 32 per GPU, 4 GPUs) |
| Learning rate | 3e-5, cosine schedule, warmup ratio 0.05 (5%) |
| Weight decay | 1e-5 |
| Seed | 42 |
| Trainable | diffusion/action head, projector and the vision-language LayerNorm, plus the top 4 layers of the language backbone — see the note below. Launch flags: --no-tune-llm --no-tune-visual --tune-projector --tune-diffusion-model |
| Augmentation | color jitter (brightness 0.3, contrast 0.4, saturation 0.5, hue 0.08) |
| Dataloader workers | 8 |
| Save cadence | every 5,000 steps, save_total_limit 2 (so checkpoint-55000 and checkpoint-60000 survived; this repo is checkpoint-60000) |
| Runtime | DeepSpeed 0.17.5, ZeRO stage 2, PyTorch 2.7.1+cu128, Transformers 4.51.3 |
All of the above is read back from the resolved experiment_cfg/conf.yaml saved next to
the checkpoint (max_steps 60000, global_batch_size 128, num_gpus 4,
learning_rate 3.0e-05, weight_decay 1.0e-05, warmup_ratio 0.05, seed 42,
save_steps 5000, the single language modality key annotation.human.task_description
and the 18 resolved dataset paths) and from this arm's submission ledger, which records
the same recipe for all three Ours target fine-tunes (deepspeed 0.17.5 ZeRO-2,
--save-steps 5000 --save-total-limit 2, effective batch 128). DeepSpeed 0.17.5 is the
pinned version for this lineage; it does not have the 0.17.6 ZeRO-2
gradient-accumulation defect, so the reported effective batch of 128 is the batch the
optimizer actually saw. The cosine schedule ran to its end (step 60,000 learning rate
2.2783020470207305e-14, loss 0.0193, grad-norm 0.366), and
best_model_checkpoint is None (no load-best-at-end), so checkpoint-60000 is the
final model by construction.
--no-tune-llm --no-tune-visual resolve to tune_llm: false and tune_visual: false
in the run's own model config, and the runtime's startup log accordingly prints
Tune backbone llm: False / Tune backbone visual: False. But that same config also
carries tune_top_llm_layers: 4, and the runtime acts on it: besides the
diffusion/action head, the projector and the vision-language LayerNorm
(tune_vlln: true), it marks the top 4 layers of the language backbone trainable —
layers 12-15, the four immediately below select_layer: 16, where the projector reads
the backbone — and casts them to fp32 because backbone_trainable_params_fp32: true.
On a sibling run of this campaign with the same training runtime and the same launch
flags, that is 160 trainable parameter tensors, every one of them under
model.language_model.model.layers.{12,13,14,15}, and the runtime reports
Trainable parameters: 1,619,968,000 (49.29%). No vision-tower parameter appears in
that list, so the vision tower is genuinely frozen.
So the accurate statement is: the vision tower is frozen and the language backbone is frozen except for its top 4 layers, which are fine-tuned along with the projector, the vision-language LayerNorm and the diffusion/action head. This is the stock behavior of this runtime at these flags and is identical across the sibling checkpoints of this campaign, whose cards now state it in the same terms.
None of this has to be taken on trust: it is readable from this repository's own
config.json, which ships tune_llm: false, tune_visual: false,
tune_top_llm_layers: 4, select_layer: 16, tune_vlln: true,
tune_projector: true, tune_diffusion_model: true and
backbone_trainable_params_fp32: true.
ours_segment_film is an arm name, not an architecture changeThe run is labelled ours_segment_film in this campaign's ledgers (arm: ours). That
is the arm name, not a claim about a modified network: the deployed training runtime
is the same tree used by the sibling arms, and the network this repository ships is
stock GR00T N1.6 at the flags listed above. The Ours-specific part of the method lives
upstream, in how the mid-training parent's subtask segmentation was produced; this
fine-tuning stage is full-task-only and therefore consumes no segmentation at all.
The publishing lab evaluated exactly this export on its own RoboCasa365 target benchmark: all 18 atomic-seen tasks × 50 episodes = 900 scored episodes, seed 0, conditioned on each environment's own full task instruction (no high-level planner in the loop). Mean per-task success rate: 0.771. This is a single-seed number from one evaluation protocol on one simulator build; it is reported for provenance, not as a benchmark claim, and no safety-performance claim is made by this model card.
| Repo | Role |
|---|---|
gr00t-n1.6-ours-racer-mt80k | the mid-training parent this run initializes from (job 47) |
gr00t-n1.6-ours-racer-ft-cs-sub60k | the subtask-conditioned low-level policy of the same arm on the composite-seen 16 split (needs a high-level planner) |
gr00t-n1.6-uvd-racer-mt80k-mix50 | UVD-RACER mid-training arm (job 12916) |
gr00t-n1.6-rdd-racer-mt80k-mix50 | RDD-RACER mid-training arm (job 13) |
gr00t-n1.6-uvd-racer-ft-at60k-fulltask | UVD-RACER full-task FT, 18 atomic-seen tasks |
gr00t-n1.6-rdd-racer-ft-at60k | RDD-RACER full-task FT, 18 atomic-seen tasks |
gr00t-n1.6-uvd-racer-ft-cs60k-fulltask | UVD-RACER full-task FT, 16 composite-seen tasks |
gr00t-n1.6-uvd-racer-ft-cu60k | UVD-RACER full-task FT, 16 composite-unseen tasks |
gr00t-n1.6-rdd-racer-ft-cu60k | RDD-RACER full-task FT, 16 composite-unseen tasks |
gr00t-n1.6-uvd-racer-pt120k-mix50 | longer-schedule UVD pretraining/mid-training sibling |
qwen3-vl-4b-uvd-racer-hl6k-cs16-final | the UVD arm's high-level subtask planner, composite-seen 16 |
qwen3-vl-4b-uvd-racer-hl6k-cu16-final | the UVD arm's high-level subtask planner, composite-unseen 16 |
gr00t-n1.6-mt60k-ds0175 | earlier mid-training checkpoint of the same lineage |
Published in the same batch: gr00t-n1.6-ours-racer-ft-at60k,
gr00t-n1.6-ours-racer-ft-cs60k, gr00t-n1.6-ours-racer-ft-cu60k and
gr00t-n1.6-rdd-racer-ft-cs-sub60k.
config.jsonprocessor_config.jsonstatistics.json (normalization statistics)embodiment_id.jsonOptimizer, scheduler, RNG state, training logs, internal experiment configs and training datasets are not included. This is a model export for inference or for starting a new fine-tuning run, not a complete optimizer-state resume bundle.
from huggingface_hub import snapshot_download
checkpoint_dir = snapshot_download(
"happyhappy-jun/gr00t-n1.6-ours-racer-ft-at60k",
# revision=<commit sha of the single publication commit>
)
The whole payload is submitted in one atomic commit on a freshly created
repository, so that commit's sha is the only model revision here. It cannot be
written inside this file (the text would change the commit it names); the pinned
snapshot_download(..., revision=...) snippet is recorded in the publishing
lab's publication log alongside the per-file SHA-256 values.
Use this directory with compatible Isaac-GR00T N1.6 model/processor code.
The model type is Gr00tN1d6; ordinary Transformers AutoModel loading has not
been validated. The processor references nvidia/Eagle-Block2A-2B-v2, which may
require separate model access/download. The preserved processor includes a
RoboCasa new_embodiment modality configuration; use the matching embodiment,
three-camera layout (robot0_agentview_left, robot0_agentview_right,
robot0_eye_in_hand), state/action conventions and normalization statistics.
Because this arm was fine-tuned with full-task language only, condition it with the full task instruction, not a subtask instruction.
Upload verification covers source checksum identity, Safetensors/index structure and Hub file integrity. It is not a new GPU inference or robot rollout evaluation beyond the evaluation section above.
The underlying model is subject to the NVIDIA noncommercial license linked
by the upstream model card. A complete, unmodified copy is included as
LICENSE.pdf. The license restricts use to noncommercial research
or evaluation.
Attribution: NVIDIA Isaac GR00T N1.6 and its underlying components. This repository is an independent derivative checkpoint and is not endorsed by NVIDIA.
Validate in simulation and use appropriate robot safety controls before any physical deployment. Outputs are not guaranteed to be safe or correct.