A Benchmark for Spatially Grounded Gesture Generation

arXiv cs.GR details a benchmark for spatially grounded gesture generation that tries to solve a familiar problem for conversational avatars: a gesture can look natural while still pointing at the wrong thing. The work introduces a dataset of roughly 2,000 pointing-annotated clips captured from naturalistic VR dialogue, each paired with ground-truth 3D referents.

The benchmark is designed to separate three concerns that are often blended together in evaluation: temporal alignment, spatial grounding, and perceived naturalness. That matters for anyone building NPCs, social VR characters, or embodied assistants, because a model that wins on motion realism may still fail at the core communicative task.

The paper also includes a flow-matching baseline, MM-Conv-Flow, and compares it with an independent retrieval-based system and human motion capture. One notable result is that geometric grounding can surpass human pointing without improving perceived naturalness, which is a useful warning for teams relying on generic motion metrics.

For developers, the practical takeaway is straightforward: referential gesture quality needs separate metrics, not a single “good gesture” score. If your pipeline touches animation generation, avatar interaction, or multimodal dialogue, this benchmark gives you a more realistic way to test whether gestures actually support communication.