gr00t-n17-bigenlight-all-svf-lora16-step5000

GR00T N1.7 Bigenlight SVF — step 5000

Research snapshot from joint DiT LoRA actor (rank 16, alpha 32) and inner soft-value training with a frozen DEAS-style environment Q. Initialized from all-data Bigenlight BC; batch 32 at this checkpoint.

model-step-5000.pt is a model snapshot, not a full optimizer recovery checkpoint or a native GR00T BC checkpoint directory. It cannot be passed directly to the standard BC inference server. A matching custom SVF loader and original BC processor/statistics are required. No standalone merged actor or hardware-validated deployment adapter is included.

Code snapshot: jun981015/gr00t-bigenlight, commit 258753d26cbf6d46e7e58b4b9c60cf480e6a38bb (private; access required).

Not validated for real-robot actuation. Perform observation replay and no-actuation tests, enforce joint/velocity limits and safe timeout handling, and provide a hardware emergency stop before real use. Training feature caches do not replace VLM inference for new live observations. Original base-model and dataset terms apply. This card grants no new license.