Improving Obstacle Avoidance in End-to-End Deep Learning by Incremental Training on Collisions
2026/03/02
Obstacle avoidance is an essential feature for autonomous robots. Although it is possible to train end-to-end deep learning models on this task, performance is not always competitive to specialized sensors, and the necessity for human-generated training data makes building such systems complex and costly. Here, we propose an incremental training method starting at the model described in [12]. By successively incrementally training the model on stereo images taken just before observed collisions, the collision rate can be significantly reduced after just a few iterations. Additionally, since training is stopped relatively early, computational effort is much lower than for a more traditional full training on the original and the new collision stereo images. No human-generated training data is needed and human intervention is minimal. A test with a model retrained on all data may even indicate that our method actually performs significantly better than full training.
Seewald, A.K. (2026): Improving Obstacle Avoidance in End-to-End Deep Learning by Incremental Training on Collisions. In Röning, J., & Filipe, J. (Eds.). (2026). Robotics, Computer Vision and Intelligent Systems: 6th International Conference, ROBOVIS 2026, Marbella, Spain, March 2-4, 2026, Proceedings. Springer Nature.