ST-HADP: Spatio-Temporal hierarchical attention diffusion policy for long-horizon generalizable bimanual visuomotor imitation.

Abstract

Introduction: Dual-arm robotic manipulation presents fundamental challenges in coordinating spatially shared perception and temporally extended behaviors under limited demonstration settings. Existing diffusion-based visuomotor policies rely on flat temporal horizons and globally pooled visual features, which fail to capture the structured nature of bimanual collaboration.

Methods: We propose the Spatio-Temporal Hierarchical Attention Diffusion Policy (ST-HADP), a framework that extends 3D diffusion policies through explicit spatial and temporal structuring. ST-HADP introduces a Spatial Attention Module that learns arm-specific focus over task-relevant 3D regions, enabling dynamic and coordinated spatial reasoning. It further incorporates a Temporal Abstraction Module that models action sequences across multiple timescales via hierarchical latent variables, facilitating coarse-to-fine action generation aligned with the natural progression of long-horizon tasks. These components are jointly optimized with a multi-objective loss function that integrates attention regularization and temporal consistency, promoting spatially focused and temporally smooth coordination.

Results: We evaluate ST-HADP on the RoboTwin 2.0 platform across six dual-arm robot configurations with diverse morphologies and tasks. Using only 50 automatically generated expert demonstrations, our method consistently outperforms baseline policies, achieving higher success rates with modest additional computational overhead.

Discussion: The results demonstrate that explicit spatial and temporal structuring enables effective dual-arm coordination under limited demonstration settings. ST-HADP provides a generalizable framework for bimanual manipulation, suggesting that hierarchical attention mechanisms offer a promising direction for sample-efficient learning of coordinated multi-arm behaviors.