Downloads · 30 days
0
RLobot-jun/gr00t-n17-bigenlight-50per-svf-qvgm10-iql-e08-q150k-proprio-k04-g025-noaug-step5000
gr00t-n17-bigenlight-50per-svf-qvgm10-iql-e08-q150k-proprio-k04-g025-noaug-step5000 is a machine learning model from RLobot-jun. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
BC-deduplicated inference weights. Requires RLobot-jun/gr00t-n17-bigenlight-50per-task-step10000 at revision 1704897ac6a2a93c1d1fd806925b9237977ed8c5.
Downloads · 30 days
0
Access
Public
Updated Sep 21, 2026
Repo size
130 MB
Likes
0
Public
Click a slice to open those files.
.safetensors130 MB · 97%
From the Hugging Face model README
BC-deduplicated inference weights. Requires RLobot-jun/gr00t-n17-bigenlight-50per-task-step10000 at revision 1704897ac6a2a93c1d1fd806925b9237977ed8c5.
q. prefix); construct its wrapped ActionConditionedQEnsemble using adapter_config.json, not the old twin-Q MLP.adapter/actor_lora.safetensors, adapter/env_q.safetensors, adapter/inner_critic.safetensors: online inference weights only.Apply actor LoRA onto the exact BC, never onto an already SVF-updated actor. No BC/VLM/reference/optimizer/EMA targets included. This is not a native BC-server folder or a training-resume checkpoint. Frozen actor/base equality is checked after training dtype conversion; real robot parity has not been tested for this export. Source code is distributed separately, not in this public artifact.
qvgm10 means the requested Q-VGM-style action-reinjected 10-head MLP, not an exact paper reproduction. iql-e08-q150k identifies the frozen IQL source (expectile 0.8, 150k updates). proprio means explicit proprio is included. k04-g025 describes SVF kappa/g, noaug means no data augmentation (flow/SDE noise remains), and the final step is SVF training progress, not IQL progress.
This is NOT the earlier plain twin-Q IQL or DEAS distributional critic. Environment Q uses the mean of 10 heads; inner soft value has 2 heads. The actor adapter workflow is unchanged, but environment-Q weight architecture differs. Matching private code is on q-vgm-critic in gr00t-bigenlight. See CRITIC_SVF_QUICKSTART.md there (repository access required). Do not load env_q.safetensors into the old twin-Q class.