CA-Safe DP: Integrating Coordinate Attention and Overshoot Recovery for Robotic Manipulation

Abstract

Diffusion Policy has demonstrated strong performance in long-horizon robotic manipulation by generating smooth and executable action trajectories through conditional denoising. However, practical deployment remains limited by two key challenges. First, inadequate modeling of pose and position features for grasping and placing tasks reduces generalization under randomized initial states. Second, frequent violations of physical execution constraints (e.g., joint position and torque limits) lead to joint overshoot and control oscillations during long-horizon inference. To address these issues, we propose an integrated framework that combines perception enhancement with execution-constraint enforcement. On the perception side, we introduce CA-ResNet by inserting a Coordinate Attention module into Layer 4 feature maps of ResNet, enabling direction-aware attention along both spatial axes. On the execution side, we design a training-free Overshoot Recovery Safety Layer with three stages: near-limit detection, adaptive recentering, and safety clipping. Experiments on NVIDIA Isaac Sim with a Franka Panda platform show that our method significantly outperforms DP-UNet and DP-Transformer baselines in success rate, positioning accuracy, and pose stability. Ablation studies further confirm the effectiveness of both components and their complementary gains.