NNPP: A Learning-Based Heuristic Model for Accelerating Optimal Path Planning on Uneven Terrain

TL;DR

This work introduces NNPP, a learning-based model that predicts heuristic regions for path planning on uneven terrain from a DEM-derived traversability cost map. By using a three-channel input (cost map plus Gaussian-encoded start and goal) and a STDC-based encoder–decoder with attention and pyramid pooling, NNPP outputs a probabilistic region for the optimal path, which guides A* to search only within that region. Across 256×256 lunar DEMs and dynamic scenarios, NNPP achieves around a 3× speedup for A* (and ~2.5× for D* variants) with minimal degradation in path quality, while delivering CPU-friendly inference at about 17 FPS. The method demonstrates strong scalability and applicability to dynamic environments, with future work aimed at refining regional predictions and achieving end-to-end path reconstruction.

Abstract

Intelligent autonomous path planning is essential for enhancing the exploration efficiency of mobile robots operating in uneven terrains like planetary surfaces and off-road environments.In this paper, we propose the NNPP model for computing the heuristic region, enabling foundation algorithms like Astar to find the optimal path solely within this reduced search space, effectively decreasing the search time. The NNPP model learns semantic information about start and goal locations, as well as map representations, from numerous pre-annotated optimal path demonstrations, and produces a probabilistic distribution over each pixel representing the likelihood of it belonging to an optimal path on the map. More specifically, the paper computes the traversal cost for each grid cell from the slope, roughness and elevation difference obtained from the digital elevation model. Subsequently, the start and goal locations are encoded using a Gaussian distribution and different location encoding parameters are analyzed for their effect on model performance. After training, the NNPP model is able to \textcolor{revision}{accelerate} path planning on novel maps.

Figures (16)

• Figure 1: In the proposed framework, a cost map with start and goal points is fed into the NNPP. The model generates a probability map where each pixel's probability value indicates the likelihood of the optimal path passing through that pixel. This probability map is then utilized to guide the ${A^\star}$ algorithm on the original map, resulting in the final optimal path.
• Figure 2: The traversal cost for each pixel was calculated using the following process: ${(a)}$ the original DEM was obtained; ${(b)}$ the slope cost was computed; ${(c)}$ the roughness cost was calculated; ${(d)}$ the elevation difference cost was determined. The resulting images, displayed from left to right, represent the original DEM, slope cost, roughness cost, and elevation difference cost. In this paper, we set ${\varphi_s = 30°}$, ${H_s = 0.2}$, ${r_s = 0.6}$.
• Figure 3: Path planning task, including a traversability cost map illustration and a path illustration
• Figure 4: The model architecture diagram. It consists of an encoder, a decoder, and a contextual structure. The input is a three-channel tensor, and the output is a single-channel tensor.
• Figure 5: Attention fusing model. The model fuses the low-level features from earlier stages with the high-level features, in order to leverage spatial details that can help produce more accurate path planning outputs.