The ability to reliably estimate free space is an essential requirement for efficient and safe robot navigation. This paper presents a novel system, built upon a stochastic framework, which estimates free space quickly from stereo data, using self-supervised learning. The system relies on geometric data in the close range of the robot to train a second-stage appearance-based classifier for long range areas in a scene. Experiments are conducted on board an unmanned ground vehicle, and the results demonstrate the advantages of the proposed technique over other self-supervised systems.

Many applications require unmanned ground vehicles (UGVs) to travel at high speeds on sloped, natural terrain. In this paper, a potential field-based method is proposed for UGV navigation in such scenarios. In the proposed approach, a potential field is generated in the two-dimensional “trajectory space” of the UGV path curvature and longitudinal velocity. In contrast to traditional potential field methods, dynamic constraints and the effect of changing terrain conditions can be easily expressed in the proposed framework. A maneuver is chosen within a set of performance bounds, based on the local potential field gradient. It is shown that the proposed method is subject to local maxima problems, rather than local minima. A simple randomization technique is proposed to address this problem. Simulation and experimental results show that the proposed method can successfully navigate a small UGV between predefined waypoints at speeds up to 7.0 m/s, while avoiding static hazards. Further, vehicle curvature and velocity are controlled during vehicle motion to avoid rollover and excessive side slip. The method is computationally efficient, and thus suitable for onboard real-time implementation.