Hierarchical Multi-Array Tactile Representation Learning and Diffusion Policy for Real-World Dexterous Placement Tasks

In contact-rich dexterous manipulation tasks, tactile feedback is critical for achieving high-precision alignment and stable physical interaction. However, signals from commonly used distributed tactile sensors are often high-dimensional and sparse, with scattered and non-uniform spatial layouts, which makes it challenging to learn generalizable tactile representations from raw measurements and to leverage them effectively for policy learning. To address these challenges, we propose a hierarchical tactile representation learning framework for real-world dexterous placement. In the local pretraining stage, we reweight zero and non-zero samples in the reconstruction loss to encourage the model to focus on physically meaningful contact regions. In the global representation stage, we further develop a multi-array tactile modeling network that integrates hand-structure priors and incorporates sensor pose information to enable structured cross-array alignment and aggregation, thereby learning stable and consistent global tactile features. For policy learning, visual observations and hierarchical tactile representations are jointly used as conditioning inputs, and a diffusion policy is adopted to model the action distribution. Experimental results demonstrate that the proposed hierarchical tactile representation learning substantially improves the success rate and robustness of policy learning, while simultaneously reducing the number of closedloop action steps required to complete a placement episode, thus enhancing overall execution efficiency.