ARMS: Anchor–Relational Motion Streaming for Seamless Solo-Social Motion Transitions

Generating temporally continuous and socially coherent human motion from natural language remains a fundamental challenge, especially in realistic streams where people act alone, enter interactions, and later disengage. Most existing text-to-motion methods generate fixed-length clips under static agent configurations. These approaches struggle with solo-social transitions and cannot easily support incremental generation over long horizons without causing boundary discontinuities or spatial drift. This paper proposes ARMS (Anchor-Relational Motion Streaming) to unify solo motion and human-human interaction within a single causal generative framework that supports seamless transitions under changing social configurations.

To balance the trade-offs between incremental temporal stability and inter-person spatial alignment, ARMS introduces a dynamics-asymmetric representation. It separates per-person temporal evolution from inter-person alignment using a partner-referenced relative-translation term. One agent is designated as the Anchor Agent (modeling canonical temporal dynamics independently of spatial relations), while the second agent is designated as the Relational Agent (storing root translation offset with respect to the Anchor). For solo motion or the Anchor stream, the relative translation term is set to zero. This design avoids global-coordinate drift and preserves stable inter-person geometry whenever interactions emerge.

ARMS compresses motion sequences into a temporally causal latent space using a shared temporal variational autoencoder. Building upon an autoregressive diffusion backbone, the model processes a concatenated latent stream of Anchor and Relational tokens through a Causal Relational Diffusion Transformer. It utilizes segment-wise causal noising and a mode-aware relational gating mask to control cross-agent visibility. In interaction mode, cross-agent connections are active, whereas in solo mode, relational tokens are masked out. This unified architecture enables the model to handle both solo actions and two-person interactions within the same streaming process.