gr00t-n17-bigenlight-50per-task-svf-iql-e09-k04-g025-step10000

N1.7 50/task SVF, IQL expectile 0.9 — step 10000

BC-deduplicated inference bundle. Load base RLobot-jun/gr00t-n17-bigenlight-50per-task-step10000 at revision 1704897ac6a2a93c1d1fd806925b9237977ed8c5 first.

• Frozen env critic: no-encoder scalar twin IQL Q, expectile 0.9, 30,000 IQL updates, 50 episodes/task, step-cost reward, gamma 0.99.
• SVF actor and inner critic checkpoint: 10000 updates, kappa=0.4, g=0.25 (c=0.64), batch 32, flow steps 4, candidates 8.
• Actor LoRA rank 16, alpha 32; two inner soft-value heads; action horizon 16.
• adapter/actor_lora.safetensors: actor LoRA only.
• adapter/env_q.safetensors: frozen online Q1/Q2.
• adapter/inner_critic.safetensors: trained online inner soft-value critics.
• adapter/adapter_config.json: architecture and provenance.

Root processor/statistics files are copied from the common BC. BC/VLM weights, reference, optimizer and target networks are not included. Apply LoRA to the matching BC ProjectedFlowActor, not an already SVF-finetuned actor. See adapter/README.md. This is not a standalone native BC model directory or training-resume checkpoint. Frozen actor weights were verified against local BC after dtype conversion; robot inference parity was not tested for this export.