Downloads · 30 days
12
44% of all-time downloads
RLobot-jun/gr00t-n17-bigenlight-50per-task-15hz-from-nvidia-bc-head-only-step20000
gr00t-n17-bigenlight-50per-task-15hz-from-nvidia-bc-head-only-step20000 is a robotics model from RLobot-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.
15Hz ONLY, not 30Hz. Independently initialized from NVIDIA GR00T-N1.7-3B and trained for 20,000 BC updates on 15Hz data. This is NOT initialized from the historical 30Hz BC or from the other fresh BC checkpoint. No SV…
Downloads · 30 days
12
44% of all-time downloads
All-time downloads
27
Public
Repo size
6.5 GB
Likes
0
Public
Click a slice to open those files.
.safetensors6.5 GB · 100%
From the Hugging Face model README
15Hz ONLY, not 30Hz. Independently initialized from NVIDIA GR00T-N1.7-3B and trained for 20,000 BC updates on 15Hz data. This is NOT initialized from the historical 30Hz BC or from the other fresh BC checkpoint. No SVF or IQL updates, critic, or LoRA adapters are included.
Data: 4 tasks × 50 episodes = 200 demonstrations (carrot, bowl stack, triple bowl stack, cube stack). Aligned image/state/action decimation: 30Hz → 15Hz. Batch 32, seed 42, AdamW LR 1e-4, cosine over 20,000 steps, 5% warmup (1000 steps), BF16 mixed precision. LLM/vision frozen; observation/state/action projections, VL LayerNorm/self-attention and full DiT/action head tuned. Standard BC ColorJitter retained; state dropout 0.
Actual action chunk: 16 × 7, absolute six arm-joint targets plus gripper.
Includes robot joint proprioception. Processor/model padding: 40 × 132.
16 actions at 15Hz cover approximately 1.067 seconds; default flow inference
uses 4 steps. Embodiment tag: NEW_EMBODIMENT; dataset robot: ur7e_gello.
VLM weights are NOT included. All trained action_head.* weights ARE included:
DiT, post-VLM LayerNorm and self-attention, state/action encoders, action decoder,
and position embedding. This is a full BC head (~6.48 GB FP32), NOT a LoRA adapter.
Processor, normalization statistics and embodiment mapping are included. Use
these files together; do not substitute older 30Hz normalization.
Optimizer, scheduler, RNG and training logs are excluded. Fresh 10k and 20k
runs have different cosine schedules.
Required shared base: nvidia/GR00T-N1.7-3B at revision 2fc962b973bccdd5d8ce4f67cc63b264d6886495. The exporter checked every omitted backbone tensor against this exact base and verified that every exported action-head tensor is bit-exact.
This repository is not directly loadable with AutoModel.from_pretrained. Download this bundle, then run its supplied CPU-only helper (torch, safetensors, and huggingface_hub are required; use your GR00T N1.7 environment):
hf download RLobot-jun/gr00t-n17-bigenlight-50per-task-15hz-from-nvidia-bc-head-only-step20000 --local-dir ./bc-head-20000
python ./bc-head-20000/restore_bc.py --output ./bc-full-20000
The helper automatically downloads the pinned NVIDIA base to the Hugging Face cache and reconstructs a full local checkpoint. If you already have that base:
python ./bc-head-20000/restore_bc.py --base-dir /path/to/GR00T-N1.7-3B --output ./bc-full-20000
Point the existing GR00T N1.7 policy/server model_path at ./bc-full-20000
and use embodiment NEW_EMBODIMENT. No actor architecture or runtime policy
changes are required. Keep bc-head-20000: restored head shards are symlinks.
The script checks the base backbone against per-tensor hashes before restoring;
the reconstructed backbone preserves the BC checkpoint dtype as well as values.
The existing GR00T constructor also downloads/loads nvidia/Cosmos-Reason2-2B.
Its normal HF access requirements still apply. That construction step is NOT a
replacement for the pinned GR00T base: the outer checkpoint restores the GR00T
backbone. Do not substitute arbitrary Cosmos/GR00T revisions for the verified base.
Code: https://github.com/jun981015/gr00t-bigenlight/tree/q-vgm-critic Use the N1.7 policy/server with the RESTORED directory and existing Bigenlight modality configuration. Check control frequency and joint/gripper conventions before real-robot deployment; no deployment safety or success guarantee.
Derived from NVIDIA GR00T-N1.7-3B. The upstream LICENSE and original model card are preserved in this repository. See LICENSE for the applicable terms.