Downloads · 30 days
9
35% of all-time downloads
RLobot-jun/gr00t-n17-bigenlight-50per-15hz-freshbc10k-svf-iql50k-e08-k06-g025-lora16-step10000
gr00t-n17-bigenlight-50per-15hz-freshbc10k-svf-iql50k-e08-k06-g025-lora16-step10000 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.
Lightweight adapter bundle: NO BC/VLM weights. 15Hz ONLY, not 30Hz. 50 episodes/task; independent fresh NVIDIA-initialized 15Hz BC10k, then 10000 SVF updates. Not the historical 30Hz BC10k - 15Hz BC10k pipeline.
Downloads · 30 days
9
35% of all-time downloads
All-time downloads
26
Public
Repo size
130 MB
Likes
0
Public
Click a slice to open those files.
.safetensors130 MB · 97%
From the Hugging Face model README
Lightweight adapter bundle: NO BC/VLM weights. 15Hz ONLY, not 30Hz. 50 episodes/task; independent fresh NVIDIA-initialized 15Hz BC10k, then 10000 SVF updates. Not the historical 30Hz BC10k -> 15Hz BC10k pipeline.
adapter/actor_lora.safetensors: DiT LoRA rank16, alpha32 (~26.2 MB).adapter/env_q.safetensors: frozen IQL50k Q ensemble (~66.8 MB).adapter/inner_critic.safetensors: trained twin soft-value heads (~37.1 MB).adapter/adapter_config.json.No duplicated BC, VLM, reference, EMA targets, optimizer, RNG, or IQL V network. Actor-only deployment needs the LoRA file and exact BC; the two critic files are optional for scoring or explicit soft-value guidance.
Environment Q: IQL expectile 0.8, 50,000 updates, scalar MSE, 10 independent action-reinjected Q heads, mean aggregation, proprio included. Q stays frozen throughout SVF. This is NOT a DEAS/HL-Gauss critic. Feature width 2212: pooled post-BC VL features + normalized proprioception + embodiment; Q also receives the clean action chunk. Valid action horizon16 × 7; model padding40 × 132. IQL action indices select the valid 112 coordinates. Reward -1 per 15Hz primitive action, successful goal0, gamma0.99 with chunk-discounted return and gamma^16 bootstrap.
SVF kappa=0.6, g=.25, c=1.44; batch32, actor/inner LR3e-4, flow4, candidates8 (batched with shared first reference velocity). Two full inner MLP heads, Fourier time16; min for regression, mean guidance gradient. VLM, BC observation projection, state/action encoder/decoder and reference frozen. No image augmentation. Existing zero-masked SVF padding.
Required exact base: RLobot-jun/gr00t-n17-bigenlight-50per-task-15hz-from-nvidia-bc-head-only-step10000, revision 970ca589335032e0fb37961dfa5caa83a5c511be. It is itself a head-only BC bundle, so restore it first:
hf download RLobot-jun/gr00t-n17-bigenlight-50per-task-15hz-from-nvidia-bc-head-only-step10000 --revision 970ca589335032e0fb37961dfa5caa83a5c511be --local-dir ./fresh-bc10k-head
python ./fresh-bc10k-head/restore_bc.py --output ./fresh-bc10k-full
restore_bc.py downloads the exact NVIDIA N1.7 base when necessary; pass
--base-dir /path/to/GR00T-N1.7-3B to reuse an existing base. Keep the source
head bundle because the restored model links its shards. Normal Cosmos HF
access is still needed by the GR00T model constructor.
Load fresh-bc10k-full using the N1.7 code and its processor/statistics.
Apply inject_dit_lora(head, rank=16, alpha=32), wrap in ProjectedFlowActor,
and call load_actor_adapter(actor, path_to_actor_lora) from
gr00t.rl.export_adapter. Target names are listed in adapter_config.json.
Do not apply this adapter to the 20k BC, another BC revision, or an already
SVF-adapted model. For env Q, use FrozenActionIQLQ's q. state prefix; the
inner network is the separate Fourier-time-conditioned FeatureCritic.
This is NOT a standalone from_pretrained model directory or a training-resume
checkpoint. Export verifies that frozen actor weights match the exact BC and
that all exported tensors are finite; no real-robot parity test is claimed.
Code: https://github.com/jun981015/gr00t-bigenlight/tree/q-vgm-critic