Abstract
Autonomous navigation in robotics relies heavily on machine learning techniques to learn from expert demonstrations and sensor data. However, traditional methods often struggle to adapt to novel or challenging environments, leading to suboptimal performance and potential safety risks. This thesis proposes a novel Bayesian approach to model terrain preferences for autonomous navigation, employing a combination of Markov Chain Monte Carlo and the Bradley-Terry model. By incorporating uncertainty into the reward weights of trajectories, our method improves the robot’s ability to navigate safely in dynamic environments. Experimental results demonstrate that our approach outperforms existing methods in terms of risk avoidance and adaptability, highlighting its potential for real-world applications.