gr00t-n17-bigenlight-all-svf-lora16-step10000

GR00T N1.7 Bigenlight SVF — step 10000

Completed 10,000-update joint DiT LoRA actor (rank 16, alpha 32) and inner soft-value run using frozen DEAS-style environment Q. All-data Bigenlight BC initialization; final batch size 32. Resumed from step 9611 to finish this run.

This model snapshot excludes the optimizer. It is not a native BC checkpoint directory and cannot be loaded directly by the standard GR00T BC server. Matching custom SVF actor/LoRA loading and the original BC processor, statistics and frozen backbone are required. No standalone robot-ready adapter is included. Code snapshot: jun981015/gr00t-bigenlight at 258753d26cbf6d46e7e58b4b9c60cf480e6a38bb (private; access required).

This completed run predates the shared-lambda/inner-target SDE optimization. It used independent SDE draws for lambda estimation and inner targets. The subsequent training-loop optimization does not alter these saved weights or the actor-only ODE sampling code.

Not validated for real actuation. Validate observation replay, normalization, joint/gripper mapping, safe limits, timeout handling and emergency stop first. Original base-model and dataset terms apply; this card grants no new license.