DistFormer+: A robotic arm collision distance estimator for multi-obstacle environments

Learning-based collision detectors have been widely applied in path planning due to their faster detection speed compared to traditional computational geometry-based collision detectors. Existing learning-based methods not only support batch detection but also estimate collision distance. These methods exhibit high accuracy in static environments with fewer obstacles. However, when faced with dynamic environments containing numerous obstacles, the accuracy of these learning-based methods declines sharply due to the complexity and diversity of the working environment. This decline in accuracy makes it difficult to support the path planner in determining a safe path, thus limiting their application scenarios. In this paper, we propose a collision detection estimator with high detection accuracy in environments with multiple moving obstacles, termed DistFormer+. This estimator introduces a novel Top-K strategy that effectively reduces the sequence length of the robotic arm while preserving the integrity of the information, and leverages the cross-attention mechanism to realize the information interaction between the robotic arm and each obstacle in parallel, obtaining the local collision distance information. Subsequently, an attention-based weighted summation selection mechanism is used to identify key information related to the global collision distance from all local collision distance information, thereby achieving accurate calculation of the global collision distance. Extensive experimental results demonstrate that the proposed estimator significantly outperforms state-of-the-art models in both estimation accuracy and inference speed in static and dynamic multi-obstacle environments.