Autonomous Mapless Navigation on Uneven Terrains

TL;DR

The paper tackles autonomous navigation on uneven terrains without relying on a global map by leveraging a Sparse Gaussian Process–based local perception model trained on LiDAR data to infer terrain elevation and uncertainty. A safety-aware cost function uses GP uncertainty to select a feasible local subgoal that keeps roll and pitch within defined bounds, enabling real-time mapless guidance toward a goal. The approach is validated in Gazebo simulations against a map-based baseline and demonstrated on a real robot, showing safer trajectories with lower orientation and elevation changes while maintaining competitive path efficiency. This work provides a practical, computationally light framework for uncertainty-aware, mapless terrain navigation with potential for real-time deployment in unstructured outdoor environments.

Abstract

We propose a new method for autonomous navigation in uneven terrains by utilizing a sparse Gaussian Process (SGP) based local perception model. The SGP local perception model is trained on local ranging observation (pointcloud) to learn the terrain elevation profile and extract the feasible navigation subgoals around the robot. Subsequently, a cost function, which prioritizes the safety of the robot in terms of keeping the robot's roll and pitch angles bounded within a specified range, is used to select a safety-aware subgoal that leads the robot to its final destination. The algorithm is designed to run in real-time and is intensively evaluated in simulation and real world experiments. The results compellingly demonstrate that our proposed algorithm consistently navigates uneven terrains with high efficiency and surpasses the performance of other planners. The code and video can be found here: https://rb.gy/3ov2r8

Figures (7)

• Figure 1: (a) Husky robot in a simulated hilly terrain environment; (b) Original (inner) vs predicted (outer) occupancy surfaces, where warmer colors indicate less occupancy; (c) Original occupancy (inner) vs. variance (outer) surface; (d) The full variance surface with subgoals shown as colored circles; (e) The original occupancy surface with the extracted subgoals.
• Figure 2: Unfolded variance surface transformed into a 2D image, where dark areas indicate the location of the observed points, and the white areas above the dark area represent the free space. The horizontal segments $\textbf{S}\{s_i\}_{i=1}^Q$ reflect the discretized terrain's elevation profile of the scene shown in Fig. \ref{['fig_sgp_oc']} as seen from the robot's sensor. $Z^{R}_{s_i}$ is the height of a segment $s_i$ in the robot's frame $\mathcal{R}$. A set of subgoals $\mathcal{G}=\{g_i\}_{i=1}^K$ is placed at the centers of the corresponding segments whose width satisfies $w_{s_i}\!<\!w_r + \delta$. The subgoals' colors indicate the cost assigned to each subgoal $g_i$ by the cost function $\textbf{J}$, where warmer colored subgoals have less cost.
• Figure 3: This graph illustrates the behavior of the steepness cost function $\mathbf{C}_{\text{stp}}$ with respect to the pitch angle $\psi$ and the relative elevation of subgoals ($dz_{g_i}$). For negative $\psi$ (indicating the robot is inclined upwards), the cost is higher when $dz_{g_i}$ is negative (subgoals are higher than the robot) compared to when $dz_{g_i}$ is positive (subgoals are lower than the robot). Conversely, for positive $\psi$ (indicating the robot is inclined downwards), the function assigns a lower cost to negative $dz_{g_i}$ values and a higher cost to positive $dz_{g_i}$ values. This structure ensures the containment of roll and pitch angles within safe limits.
• Figure 4: Path comparison in the CHT environment between our proposed approach and the baseline. The color variations represent the terrain elevation, with red shades denoting higher areas.
• Figure 5: Illustrations of roll and pitch angle variations and elevation changes during the AB path trials, as depicted in \ref{['fig:scenario1']}.