MotionVLA Uses Dual-Codebook DCT Quantization for Humanoid Motion Generation

Single-codebook motion tokenization misrepresents high-frequency velocity components — five DCT coefficients capture 93% of joint-position energy but only 37% of joint-velocity energy. MotionVLA addresses this with frequency-domain analysis motivating a dual-codebook approach to separately quantize pose and velocity signals in an autoregressive VLA.