ReactionMamba: Generating Short & Long Human Reaction Sequences

We present ReactionMamba, a framework for generating short and long 3D human reaction motions. ReactionMamba integrates a motion variational autoencoder with Mamba-based state-space models to decode temporally consistent reactions. The model generates reaction sequences conditioned on the observed actor motion and the initial pose of the reactor.

We evaluate ReactionMamba on NTU120-AS, Lindy Hop, and InterX, demonstrating competitive performance in realism, diversity, and long-sequence generation, while achieving substantial improvements in inference speed.

ReactionMamba is a conditional VAE composed of a Mamba-based encoder, a conditioning module, and a Mamba-based decoder. During training, the ground-truth reaction is encoded into a latent representation. The decoder reconstructs the reaction from the latent code, the actor motion sequence, and the initial reactor pose. At inference time, the encoder is removed and latent variables are sampled to generate diverse reactions.

We compare ReactionMamba against ground truth and representative baselines on short and long human reaction sequences. Each row shows the same interaction using Ground Truth, ReactionMamba, and one competing method.

This project is supported by the France 2030 program – ANR-22-EXEN-0004 and by the National Research Center for Applied Cybersecurity ATHENE.