Reproduce “Imitating Human Behaviour with Diffusion Models” 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
Tim Pearce, Tabish Rashid, Anssi Kanervisto, Dave Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan and 3 more
Microsoft Research demonstrates the effective use of diffusion models for imitating human behavior, learning expressive observation-to-action mappings. The method achieves an 89% task completion rate in a robotic kitchen environment, outperforming prior state-of-the-art, and yields higher game scores in Counter-Strike: Global Offensive by generating more human-like action sequences.
Key insights
- Diffusion models effectively learn and reproduce the inherent stochasticity, multimodality, and complex correlations present in human behavioral data, addressing core limitations of traditional BC.
- Classifier-Free Guidance, while effective in other diffusion applications, negatively impacts observation-to-action models by biasing agents towards less common actions strongly tied to specific observations.
- Novel guidance-free sampling strategies (Diffusion-X and Diffusion-KDE) are essential for generating reliable, high-likelihood actions suitable for real-time sequential decision-making.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.