ARFlow : Human Action-Reaction Flow Matching
with Physical Guidance

Human action-reaction synthesis, a fundamental challenge in modeling online causal human interactions, plays a critical role in applications ranging from virtual reality to social robotics. While diffusion-based models have demonstrated promising performance, they exhibit two key limitations for interaction synthesis: reliance on complex noise-to-reaction generators with intricate conditional mechanisms, thus limiting to unidirectional generation, and frequent body penetrations in generated motions. To address these issues, we propose Action-Reaction Flow Matching (ARFlow), a novel paradigm that establishes direct action-to-reaction mappings, eliminating the need for complex conditional mechanisms and supporting bi-directional generation. Directly applying traditional guidance algorithms tends to undermine the quality of generated reaction motion. We analyze the sampling of flow matching in depth and reveal an issue (Initial Point Deviation) which causes the sampling trajectory to ever farther from the initial action motion. Thus, we propose a reprojection guidance method, RE-GUID, to correct this deviation to enable better interaction. To further enhance the reaction diversity, we incorporate randomness into the sampling process. Extensive experiments on NTU120, Chi3D and InterHuman datasets demonstrate that ARFlow not only outperforms existing methods in terms of FID and motion diversity but also significantly reduces body collisions, as measured by our introduced Intersection Volume and Intersection Frequency metrics.

Our proposed Human Action-Reaction Flow (ARFlow). We directly establish a mapping between the action and reaction distribution and our sampling process is further guided by our reprojection guidance method (RE-GUID). The change of colors represents the variation of the h-frame reaction with sampling timestep t.

Pipeline of ARFlow. (a) For a sampled timestep t, we linearly interpolate a coupled action-reaction pair to produce the intermediate state xt, which is then turns into a d-dimensional latent feature through a linear layer. We use Transformer Decoder Units to directly predict clean reaction motions. (b) After training the networks in (a), our ARFlow uses them for x1-prediction based sampling. The sampling process is further guided by our reprojection guidance method (RE-GUID) to prevent body penetrations.