Direct Action-Manifold Modeling for Efficient and Robust Bimanual Manipulation
Direct Action-Manifold Modeling for Efficient and Robust Bimanual Manipulation
This paper introduces the Direct Manifold Action Transformer (DMAT), a method that reformulates bimanual action generation as direct modeling of observation-conditioned joint-space action chunks to eliminate iterative inference latency. By anchoring predictions with local 3D geometry and applying State-Consistent Spectral-Time Refinement, DMAT achieves single-step inference while maintaining robustness under physical interventions. Evaluated across 567 standard and out-of-distribution trials, DMAT demonstrates superior aggregate success rates and faster generation speeds compared to Improved 3D Diffusion Policy (IDP3) and DiTFlow3D, establishing a new paradigm for efficient bimanual manipulation.
The authors propose DMAT, a transformer-based model that directly models action chunks conditioned on vision and proprioception to avoid iterative denoising, reporting aggregate success rate improvements of 129.1% over Improved 3D Diffusion Policy (IDP3) and 95.7% over DiTFlow3D across 567 trials including out-of-distribution cases.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
trials
567
Reported trials
View original evidence
ctory-straightening procedures. Within DMAT, local 3D geometry anchors predictions to spatial structure, while State-Consistent Spectral-Time Refinement (STR) reconciles predicted chunks with proprioception and suppresses high-frequency residuals. Across 567 standard and out-of-distribution (OOD) trials per method, DMAT outperformed Improved 3D Diffusion Policy (IDP3) and DiTFlow3D in aggregate success rate, achievin
Open source S4speedup factor
2.06
View original evidence
g relative improvements of 129.1% and 95.7%, respectively, while generating action chunks 2.06x and 1.34x faster and remaining robust under repeated 50 N interventions. These results pave the way for a novel generative paradigm for efficient and robust bimanual manipulation.
Open source S5speedup factor
1.34
View original evidence
g relative improvements of 129.1% and 95.7%, respectively, while generating action chunks 2.06x and 1.34x faster and remaining robust under repeated 50 N interventions. These results pave the way for a novel generative paradigm for efficient and robust bimanual manipulation.
Open source S5intervention force
50
View original evidence
g relative improvements of 129.1% and 95.7%, respectively, while generating action chunks 2.06x and 1.34x faster and remaining robust under repeated 50 N interventions. These results pave the way for a novel generative paradigm for efficient and robust bimanual manipulation.
Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ctory-straightening procedures. Within DMAT, local 3D geometry anchors predictions to spatial structure, while State-Consistent Spectral-Time Refinement (STR) reconciles predicted chunks with proprioception and suppresses high-frequency residuals. Across 567 standard and out-of-distribution (OOD) trials per method, DMAT outperformed Improved 3D Diffusion Policy (IDP3) and DiTFlow3D in aggregate success rate, achievin
Open source S4
g relative improvements of 129.1% and 95.7%, respectively, while generating action chunks 2.06x and 1.34x faster and remaining robust under repeated 50 N interventions. These results pave the way for a novel generative paradigm for efficient and robust bimanual manipulation.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org