Traversability-aware Adaptive Optimization for Path Planning and Control in Mountainous Terrain

TL;DR

The paper tackles autonomous navigation in extreme mountainous terrain by introducing TAO, a Traversability-aware Adaptive Optimization framework that couples apparent traversability $M_\tau$, derived from exteroceptive depth cues, with relative traversability $Ψ_τ$, derived from exteroceptive and proprioceptive cues. It integrates these signals into both planning and control: $M_\tau$ biases sampling-based path planning and constrains path selection, while $Ψ_τ$ modulates velocity in an adaptive NMPC/MPPI controller to accommodate ground-robot interactions. The approach is validated in simulation across 27 terrain types and 108 incline-augmented variants and in real-world Husky experiments, showing improved robustness, hazard avoidance, and resilience to slip and terrain undulations compared with baselines such as DEM, $M_τ$-Only, and PUTN. The work demonstrates that a dual-traversability perspective can significantly enhance traversability in complex mountainous environments and offers a modular framework that can augment existing planning and control pipelines, with potential extensions to incorporate vision for seasonal terrain changes.

Abstract

Autonomous navigation in extreme mountainous terrains poses challenges due to the presence of mobility-stressing elements and undulating surfaces, making it particularly difficult compared to conventional off-road driving scenarios. In such environments, estimating traversability solely based on exteroceptive sensors often leads to the inability to reach the goal due to a high prevalence of non-traversable areas. In this paper, we consider traversability as a relative value that integrates the robot's internal state, such as speed and torque to exhibit resilient behavior to reach its goal successfully. We separate traversability into apparent traversability and relative traversability, then incorporate these distinctions in the optimization process of sampling-based planning and motion predictive control. Our method enables the robots to execute the desired behaviors more accurately while avoiding hazardous regions and getting stuck. Experiments conducted on simulation with 27 diverse types of mountainous terrain and real-world demonstrate the robustness of the proposed framework, with increasingly better performance observed in more complex environments.

Figures (8)

• Figure 1: Robot Navigation in Mountainous Terrain. We propose a traversability-aware navigation framework in both virtual and real environments. Virtual environments are reconstructed from actual mountainous terrain. The top right of the figure displays the optimal path (cyan) calculated using traversability information.
• Figure 2: Overall Structure. Our framework takes inputs such as a height/normal map, localization, and the robot's inertial states (orientation $\mathbf{q}$, linear velocity $\mathbf{\zeta}$, and angular velocity $\mathbf{\omega}$). The computed traversability/constraint map of the terrain is used as guidance to generate a trajectory. The controller adjusts the robot's velocity with adaptive optimization based on apparent and relative traversability. This enables the robot to navigate extreme mountainous terrains robustly. The top left of the figure shows a sample of terrains collected on campus and mountains.
• Figure 3: Visualization of Exteroceptive Features. The color of each map is represented by a range from deep blue with a score of zero to vibrant yellow with a score of one. (d) is obtained as a linear combination of the components (a-c).
• Figure 4: Navigation in Mountainous Environments. We compare 4 different path planning methods - DEM (purple), $\mathcal{M}_{\tau}$-Only (green), PUTN (cyan) and TAO-Plan (red) - using Adaptive MPPI controller. The goal point is indicated by a yellow marker. Our proposed method (TAO) successfully navigates in mountainous terrains by computing trajectories that circumvent bumpy and risky regions and dynamically adjust the robot's velocity.
• Figure 5: Path Generation Process. The difference between (a) and (b) shows the sampling bias. (c) and (d) also visualizes the search tree and optimal path overlaid on the traversability map. The search tree is depicted as green solid lines, while the selected optimal path is shown as red solid lines.