Per-dimension CDF normalizer using piecewise-linear empirical CDF. - Normalize: y = 2 * CDF(x) - 1 in [-1, 1] - Unnormalize: x = CDF^{-1}((y + 1)

Configuration for Gemma model variants.

ResnetBlock with grouped convolutions for branch isolation.

ResnetBlock with grouped convolutions for branch isolation.

Track a series of values and provide access to smoothed values over a window or the global series average.

@func: Automatic Mixed Precision, AMP @info: Mixed precision training refers to using FP32 as the sovereign weight, using FP16

Processor step to prepare the state and tokenize the language input.

Core MINT PyTorch model.

@func:

PaliGemma model with action expert for MINT.

Split 7-dim action into 3 modality branches and embed each to D_embed dimensions. Input: (B, T, 7) where 7 = [pos(3), rot(3), grip(1)] Ou

Split 7-dim action into 3 modality branches and embed each to D_embed dimensions. Input: (B, T, 7) where 7 = [pos(3), rot(3), grip(1)] Ou

Mirror of PatchwiseEmbedding - project 3*D_embed back to 7-dim action. Input: (B, 3*D_embed, T) from decoder Output: (B, T, 8) -> [pos(3)

Mirror of PatchwiseEmbedding - project 3*D_embed back to 7-dim action. Input: (B, 3*D_embed, T) from decoder Output: (B, T, 8) -> [pos(3)

MINT Policy for LeRobot.