Differentiable Obstacle Avoidance Framework for Robot Safety
PUBLICATION DATE: 2 February, 2026 PUBLICATION AUTHOR/S: Louis Fernandez; Sheila Sutjipto; Victor Hernandez Moreno; Anh Minh Tu; Marc G. CarmichaelObstacle avoidance is a fundamental requirement for safe and reliable robot operation. Due to the lack of techniques capable of efficiently and accurately computing volumetric distances and their derivatives, many obstacle avoidance frameworks neglect the volumetric characteristics of objects by using points or spheres as the choice of object representation. To overcome this, an optimisation-based method is used in this letter to accurately compute the minimum distance between convex shapes. In order to enable integration with gradient-based obstacle avoidance algorithms, a method that calculates the gradient of the minimum distance is introduced. This information is incorporated into a Quadratic Program to demonstrate the effectiveness of the proposed approach for online obstacle avoidance. In contrast to methods that rely on closed-form expressions, the proposed framework leverages optimisation to simultaneously compute pairwise distances between multiple convex shapes. Simulation results, averaged over 1000 runs, demonstrate that the proposed approach completes tasks more accurately and consistently, improving trajectory tracking accuracy by up to 35% compared to the state-of-the-art baseline. The proposed framework is further validated on a physical robot system, confirming its practical applicability and robustness on real hardware.
RELATED PROGRAM/S:Human-Robot-Interaction RELATED PROJECT/S:
Project 2.3: Safe and Efficient Human-Robot Collaboration through Predictive Obstacle Avoidance in Dynamic Environments Publication link
View all publications