Event

DMAT: Single-Step Inference Directly Models Bimanual Action Manifolds

1 reports · 1 sources

Hongren Wang and co-authors propose the Direct Manifold Action Transformer (DMAT), which reframes bimanual action generation as directly modeling observation-conditioned clean joint-space action chunks from random interpolation, enabling single-step inference without reliance on iterative denoising or auxiliary steps such as trajectory straightening. The authors report that across 567 standard and out-of-distribution (OOD) trials per method, DMAT achieved higher overall success rates than Improved 3D Diffusion Policy (IDP3) and DiTFlow3D, with relative improvements of 129.1% and 95.7%, respectively, and action chunk generation speeds 2.06x and 1.34x those of the latter two, while remaining robust under repeated 50 N external force interventions. These comparison and robustness conclusions come from the authors' self-reported experimental results.