Beyond MoCap: Scaling Motion Tokenizers with Synthetic Human Motion for Generative Modeling

Researchers have proposed two frameworks for improving human motion prediction and generation, addressing limitations in existing motion capture datasets and model architectures.

A team of researchers has introduced a framework for expanding the motion representation space by leveraging large-scale synthetic human motion^[1]. Existing motion capture datasets are limited in their diversity, predominantly containing common, repetitive actions and failing to cover complex human movements^[1]. The proposed framework uses a data generation pipeline to produce diverse, physically plausible motion sequences beyond the distribution of existing datasets. It also integrates with a redesigned VQ-VAE tokenizer that adapts to this expanded motion space, jointly scaling both the training distribution and the discrete codebook. This approach enables the model to capture a richer set of motion primitives, improving the coverage and compositionality of the learned motion vocabulary. In a related development, another research team has proposed a non-autoregressive transformer for predicting long-term human motion, using spatio-temporal attention and able to process varying lengths of history observations^[2]. Human motion prediction is relevant in various domains, including human-robot interaction and autonomous driving. The proposed model addresses limitations in existing architectures, such as bidirectional attention mechanisms that can accumulate errors over time^[2].