Reproduce “DartControl: A Diffusion-Based Autoregressive Motion Model for Real-Time Text-Driven Motion Control” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kaifeng Zhao, Gen Li, Siyu Tang

A diffusion-based autoregressive motion model, DARTCONTROL, enables real-time generation of long, continuous human motion sequences from text while integrating precise spatial control. It demonstrates superior motion realism (FID 3.79), a 10x speedup over baselines (334 frames/second), and achieves 100% success in goal-reaching tasks with improved physical plausibility.

Key insights

  • Modeling human motion as a sequence of overlapping, canonicalized motion primitives simplifies the generative task and enables efficient, real-time autoregressive generation.
  • Compressing raw motion primitives into a compact, clean latent space via a VAE mitigates artifacts and significantly improves the quality and efficiency of the diffusion process.
  • The learned latent motion space allows for versatile spatial control, enabling both gradient-based optimization and efficient Reinforcement Learning for precise geometric tasks alongside semantic control.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.