Per-link collision depth prediction for redundant manipulators in operational environments

Abstract

The subject matter of this study is a collision checking for redundant robotic manipulators operating in variable environments, which remains a significant computational bottleneck in motion planning. The goal of this study is to improve computational efficiency of collision checking for multi-joint robotic manipulators in sampling-based motion planning, while preserving functional safety. This is achieved by developing and evaluating a learning-based method that predicts per-link penetration depth and serves as a statistical pre-filter rather than a replacement for exact collision checking.  The tasks are as follows: 1) to propose a novel input representation that fuses the manipulator's kinematic state with localized geometric context extracted from the environment via voxel grids; 2) to design and implement a hybrid neural network architecture combining a fully-connected projection layer with a Kolmogorov-Arnold Network (KAN); 3) to train the network on a large, procedurally generated dataset of diverse collision scenarios; and 4) to evaluate the model's regression accuracy, classification performance, and computational speedup over a direct physics-based checker. The following results were obtained: the trained model achieves high regression accuracy with a low Mean Squared Error of 0.000148 on the test set; the model achieves promising classification results with a per-link recall of 93.01%, which is an important indicator for its use as a pre-filter capable of screening out the majority of hazardous states; computational speedup - performance benchmarks for a batch of 8192 states show that the proposed approach, including data preparation and inference, is approximately 3.7 times faster than a direct physics-based checker. Conclusions. The scientific novelty of results obtained is as follows: 1) a neural network architecture combining fully-connected and Kolmogorov–Arnold Network layers is proposed for predicting per-link collision depth of a redundant manipulator; 2) integration of kinematic and voxel-based geometric features into a unified input representation for accurate collision estimation. The proposed method effectively serves as a pre-filter for sampling-based planners, reducing the number of expensive collisions checks and accelerating the overall motion planning process.