GR00T-N1.5-robocasa-atq-v22so3-granularity-2x4x-60k-20260923
RoboCasa ATQ v22, SO(3) merged rotation targets, 2x/4x groups, four experts (source horizons 16/8; output horizons [8,4,2,4]). Trained from the base GR00T model for 60,000 steps with global batch 64, seed 42. Pooling: mean, shared by router and confidence head; state features still use mean pooling. Router KL coefficient: 0.1. Balance coefficient: 0.05. Confidence regression remains enabled (coefficient 0.1). Both groups now compress delta actions: low-conf group uses 2x, high-conf group uses 4x. Router selects source horizon within the group. The original v22 conf labels are reused, without relabeling for 2x/4x. SO(3) rotation composition is applied directly to groups of 2 or 4 source actions; discrete actions use the group's last value. Conf input detach and all loss weights retain baseline values. Requires the included moe_fine_speed support in GR00T-action-quantization. Matching source is included under granularity_implementation/. Final inference weights and normalization metadata only; optimizer states omitted. Not yet evaluated. Full settings and weight hashes: ablation.json.