Downloads · 30 days
13
36% of all-time downloads
happyhappy-jun/gr00t-n1.6-rdd-racer-ft-cs-sub60k
gr00t-n1.6-rdd-racer-ft-cs-sub60k 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. It is not an official NVIDIA release. This repository is released anonymously for ICLR 2027 review.
Downloads · 30 days
13
36% of all-time downloads
All-time downloads
36
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. It is not an official NVIDIA release. This repository is released anonymously for ICLR 2027 review.
Unlike the flat siblings of this campaign, this checkpoint is one half of a
hierarchical (HL → LL) system. It is the low-level (LL) policy: it was
fine-tuned from the RDD-RACER mid-training checkpoint
happyhappy-jun/gr00t-n1.6-rdd-racer-mt80k-mix50
on the 16 composite-seen (CS) RoboCasa365 target tasks conditioned on subtask
language only, never on the full task instruction. It is the RDD-RACER twin of
gr00t-n1.6-ours-racer-ft-cs-sub60k:
same target task population, same subtask-conditioned recipe, same modality config —
only the segmentation method behind the subtask annotations, and the mid-training
parent, differ.
This LL checkpoint alone is not a complete hierarchical system. The high-level subtask planner (HL) for this arm has not been trained or published yet. The model expects to receive, at every step, a single short subtask sentence of the kind an HL planner emits — it does not decompose a full task instruction by itself, and it was never trained on one.
To run it you must supply that subtask stream from somewhere:
Conditioning this checkpoint with a full task instruction is out of distribution and should not be read as a fair evaluation of the arm.
ft_cs_sub of the rdd arm in the iclr2027_ft
fine-tuning harnesscheckpoint-60000 (60,000 optimizer updates)181084 on an internal lab Slurm cluster (4 GPUs, single node); the run
started 2026-09-14 10:57:52 KST and finished 2026-09-14 22:26:24 KST
(11 h 27 m 20 s wall, train_runtime 41,240 s), writing its COMPLETE.json marker
(global_step 60000, microbatch 32, accumulation 1, num_gpus 4,
effective_batch 128, source_job 181084, losses 0.0293 / 0.0306 / 0.0272)checkpoint-80000 published as
happyhappy-jun/gr00t-n1.6-rdd-racer-mt80k-mix50
(job 13), loaded weights-only through --base-model-path with a fresh optimizer
state — no optimizer, scheduler or RNG state carried over. The launch contract
records the SHA-256 of all seven files of that base checkpoint, and all seven match
the published MT 80K repository byte-for-byte (including
config.json cd265fe9… and model.safetensors.index.json 4baab54d…, the same two
this repo ships), so the initialization identity is verified rather than assumedd0814e7 (through the MT 80K parent)cf5148665bbce9479e1380615383d6807e6c4d4dc8afdd63389580ed3059475fThe 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 16 RoboCasa365 composite-seen (CS16) target LeRobot datasets (staged as rdd_hivla_cs16_staged), equal mix ratios (1 each); 8,077 episodes / 6,002,265 frames — the launch contract's own data_identity {tasks 16, episodes 8077, frames 6002265}. This is the same CS16 episode population the Ours-arm low-level twin saw, so at the FT stage the two arms differ only in the subtask annotations |
| Tasks | DeliverStraw, GetToastedBread, KettleBoiling, LoadDishwasher, PackIdenticalLunches, PreSoakPan, PrepareCoffee, RinseSinkBasin, ScrubCuttingBoard, SearingMeat, SetUpCuttingStation, StackBowlsCabinet, SteamInMicrowave, StirVegetables, StoreLeftoversInBowl, WashLettuce |
| Language conditioning | subtask language only. The modality config exposes the single language key annotation.human.subtask, a per-frame index into the dataset's meta/tasks.jsonl resolved by the loader. No full-task instruction field is consumed |
| Annotation source | the RDD CS16 annotation overlay (launch contract data_identity: prediction_method rdd_racer, language_source column, overlay manifest SHA-256 6d4971b9c85068ad7cc55366d4753f52fc58e16c1f9091176a488d0a13efae56), which stores the per-frame subtask sentence as a parquet column |
| Modality config | subtask_column_config.py (SHA-256 5f3602d4a45d8a813e614584e426e01b14bc2fc89588336fd24e25ca8a0332bb) — the same file, byte for byte, that the Ours-arm low-level twin used; it is the frozen runtime's RoboCasa365 config (same 3-camera video block, same 5-field state block, same 16-step action chunk) except that the language modality reads annotation.human.subtask |
| 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 run's own launch_contract.json, which records
the exact launch argument vector (--embodiment-tag NEW_EMBODIMENT --num-gpus 4 --global-batch-size 128 --gradient-accumulation-steps 1 --max-steps 60000 --learning-rate 3e-5 --weight-decay 1e-5 --warmup-ratio 0.05 --seed 42 --dataloader-num-workers 8 --save-steps 5000 --save-total-limit 2 --no-tune-llm --no-tune-visual --tune-projector --tune-diffusion-model --color-jitter-params brightness 0.3 contrast 0.4 saturation 0.5 hue 0.08) together with microbatch 32 / accumulation 1 / num_gpus 4 / effective_batch 128, language annotation.human.subtask and the 16 resolved dataset paths. The training log records
the launcher's own computed line [deepspeed] version=0.17.5 stage=2 grad_accum=1 num_gpus=4 effective_batch=128 and Loaded modality config: …/subtask_column_config.py.
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 2.278e-14, loss 0.0272, grad-norm 0.189), and
best_model_checkpoint is None (no load-best-at-end), so checkpoint-60000 is the
final model by construction.
The deployed training runtime tree is byte-for-byte the one the sibling arms used
(runtime_sha256 cdd0905e75415e5c00a383242b8d925aaf9218ba9fac40ab89f77b3cc7b820e6), so
the only recipe difference from a full-task target FT is the language key and the
annotation overlay behind it. The network this repository ships is stock GR00T N1.6 at
the flags listed above.
--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.
The RDD-RACER mid-training parent
(gr00t-n1.6-rdd-racer-mt80k-mix50)
was trained on an RDD metadata snapshot that excludes 641 episodes present in the
corresponding UVD snapshot (32,043 UVD episodes → 31,402 there). That caveat is a
property of the mid-training stage and it propagates into this checkpoint through its
initialization: an MT-stage comparison between the RDD and UVD arms is not matched in
episode population. It does not apply to this fine-tuning stage, whose 16 target
datasets carry the same 8,077-episode / 6,002,265-frame population as the Ours-arm
low-level twin.
This export had not been evaluated at publication time — no rollout evaluation of this checkpoint had been run, and evaluating it at all requires the high-level planner described above or an oracle subtask stream. No task-success or safety-performance claim is made by this model card.
| Repo | Role |
|---|---|
gr00t-n1.6-rdd-racer-mt80k-mix50 | the mid-training parent this run initializes from (job 13) |
gr00t-n1.6-ours-racer-ft-cs-sub60k | the Ours-arm low-level twin of this checkpoint: same 16 composite-seen tasks, same subtask-conditioned recipe (job 66) |
gr00t-n1.6-rdd-racer-ft-at60k | the full-task-language RDD-RACER FT on the 18 atomic-seen tasks |
gr00t-n1.6-rdd-racer-ft-cu60k | the full-task-language RDD-RACER FT on the 16 composite-unseen tasks |
gr00t-n1.6-uvd-racer-ft-cs60k-fulltask | the full-task-language comparator on the same 16 composite-seen tasks (UVD arm) |
gr00t-n1.6-ours-racer-mt80k | Ours-RACER mid-training arm (job 47) |
gr00t-n1.6-uvd-racer-mt80k-mix50 | UVD-RACER mid-training arm (job 12916) |
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 on the same composite-seen 16 split — an example of the HL half; it is not the planner for this LL checkpoint |
qwen3-vl-4b-uvd-racer-hl6k-cu16-final | the UVD arm's high-level planner on the composite-unseen 16 split |
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-rdd-racer-ft-cs-sub60k",
# 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.
Condition it with a single subtask sentence per step (see the pending-HL
section above), not with a full task 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.
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.