Active Traversability Learning via Risk-aware Information Gathering for Planetary Exploration Rovers
Traversability prediction enables safe and efficient autonomous rover operation on deformable planetary surfaces. Revealing spatial distribution from terrain geometry to rover slip behavior is key to assessing prospective traversability, but is hindered by insufficient in situ measurements on hazardous states due to conservative rover traverses. To achieve a more accurate prediction, this paper proposes a framework that actively learns latent traversability by exploring informative terrain under the constraints of stochastic rover slip. With a Gaussian process (GP) modeling the spatial distribution, we devise an iterative two-stage framework that gradually refines the model estimation, combining risk-aware informative path planning and GP updates by taking in situ measurements. The path planning stage employs our designed sampling-based algorithm to generate informative trajectories with fault-tolerant risk inference, while the GP is cautiously updated with traverse data to avoid rover immobilization. Chance constraint formulation is exploited in the framework to infer the stochastic reachability of informative regions. Through GP estimates reducing uncertainty, the algorithm incrementally reaches informative yet hazardous states along feasible trajectories. Simulation studies in rough terrain environments demonstrate that the proposed framework gathers informative traverse data while averting rover stuck situations to estimate the latent traversability model.