Downloads · 30 days
0
RLobot-jun/gr00t-n17-bigenlight-50per-task-svf-iql-k04-g025-step10000
gr00t-n17-bigenlight-50per-task-svf-iql-k04-g025-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.
Fresh SVF run from 50/task BC checkpoint10000 (200 episodes total). Frozen original scalar IQL Q: expectile0.7, 30k, WITHOUT the added pooled encoder. Trainable actor DiT LoRA rank16/alpha32 and inner critic; no env-Q…
Downloads · 30 days
0
Access
Public
Updated Sep 21, 2026
Repo size
6.8 GB
Likes
0
Public
Click a slice to open those files.
.pt6.7 GB · 98%
From the Hugging Face model README
Fresh SVF run from 50/task BC checkpoint10000 (200 episodes total). Frozen original scalar IQL Q: expectile0.7, 30k, WITHOUT the added pooled encoder. Trainable actor DiT LoRA rank16/alpha32 and inner critic; no env-Q updates. Kappa0.4, g0.25, c=kappa^2/g=0.64 throughout this run. Shared lambda/inner SDE draws, 4 integration steps, 8 candidates, no guidance norm clipping. Batch32, Adam lr3e-4, gamma0.99, reward-1 per action and0 at terminal success. Action horizon16, real16x7 actions, internal40x132 padding.
The model-only snapshot contains actor, reference, frozen env Q, inner critic and target states. No optimizer or VLM weights; IQL V is not used or included. Use matching custom code from jun981015/gr00t-bigenlight (private) and the original 50/task BC backbone/processor/statistics. This is not directly loadable by the native BC server. Training configuration is included separately. Completion is checked against the atomically committed recovery pointer. Not validated for real actuation. Original model/dataset terms apply.
Use adapter/ instead of the legacy 6.66GB model-step-10000.pt.
The common BC is RLobot-jun/gr00t-n17-bigenlight-50per-task-step10000
at revision 1704897ac6a2a93c1d1fd806925b9237977ed8c5, including processor files.
Actor LoRA is26.2MB; optional env Q and inner critic are37MB each.
Download with snapshot_download(repo_id=REPO_ID, allow_patterns=["adapter/*"])
to avoid downloading the legacy full checkpoint. For actor-only use, select
adapter/actor_lora.safetensors and adapter/adapter_config.json.
Frozen actor tensors were verified against BC after BF16 conversion; adapter
and critic tensors were checked exactly against the source SVF snapshot.
This is not a full hardware/rollout validation. Do not apply LoRA twice.
The new adapter loader/export helper is currently in the local development
checkout; code distribution is separate from this public weight upload.