cs.ROMar 18, 2026

Neural Radiance Maps for Extraterrestrial Navigation and Path Planning

Authors: Adam Dai, Shubh Gupta, Grace Gao

Organizations: Stanford University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Autonomous vehicles such as the Mars rovers currently lead the vanguard of surface exploration on extraterrestrial planets and moons. In order to accelerate the pace of exploration and science objectives, it is critical to plan safe and efficient paths for these vehicles. However, current rover autonomy is limited by a lack of global maps which can be easily constructed and stored for onboard re-planning. Recently, Neural Radiance Fields (NeRFs) have been introduced as a detailed 3D scene representation which can be trained from sparse 2D images and efficiently stored. We propose to use NeRFs to construct maps for online use in autonomous navigation, and present a planning framework which leverages the NeRF map to integrate local and global information. Our approach interpolates local cost observations across global regions using kernel ridge regression over terrain features extracted from the NeRF map, allowing the rover to re-route itself around untraversable areas discovered during online operation. We validate our approach in high-fidelity simulation and demonstrate lower cost and higher percentage success rate path planning compared to various baselines.

Figures & tables

Figure 1: Examples of global plan re-routing due to updates from local observations from the Curiosity rover.
Figure 2: Overview of the Neural Radiance Field (NeRF) training and novel view synthesis task. Given a set of 2D training images with camera poses, the NeRF trains a neural network to represent the scene. Then, given a target camera pose, the NeRF can synthesize a new image from that camera pose of the scene.
Figure 3: Inputs and outputs of the NeRF neural network. The network takes a 3D point in the scene and viewing direction of the point as input, and learns to predict the color and volumetric density of that point when viewed from that direction. In order to synthesize an image, a NeRF uses multiple rays (viewing directions), and for each ray, queries points along the ray and accumulates their predicted color and densities to compute an RGB pixel value.
Figure 4: AutoNav local costmap pipeline. First, a depth image ( 4(b) ) is extracted from stereo RGB images from the rover’s onboard cameras ( 4(a) ). Next, the depth image is projected into 3D to produce a point cloud in local frame ( 4(c) ). Finally, the 3D space is gridded up into a 2D costmap, where the cost of each cell is determined from the gradient (slope) and variance (roughness) of points within.
Figure 5: Overview of online planning framework. The rover uses a global costmap to plan a global path for navigation. The global costmap is continuously updated through onboard imagery processing, resulting in the creation of a local costmap which reflects the current terrain and obstacles around the rover. The rover effectively integrates the local costmap into the global costmap using a NeRF-based global representation, facilitating the adaptation of the local cost values to multiple regions within the global costmap. The rover then combines the new global path with the local costmap values to determine a local path for execution.
Figure 6: Neural Radiance Map construction process. First, imagery of the global region is gathered from simulated camera trajectories. Next, poses of these images are estimated, and the images and poses together are used to train a NeRF of the global region.
Figure 7: NeRF terrain feature extraction. For a given point in the scene, we render a group of rays in a surrounding patch of terrain. For each ray, we accumulate color and densities from the NeRF to obtain both color and depth values. We then concatenate all the values together into a single vector, which forms the feature vector for given point.
Figure 8: Comparative visualization of NeRF renders and original imagery for both simulated Mars GES data and Unreal Engine Moon environment. The original training images are visualized alongside their NeRF renders, which possess a high degree of similarity to the original images.
Figure 9: Moon simulation environment. The environment features realistic lighting and contains highly detailed 4×44\times 4 km with craters and rocks.
Figure 10: Screenshot of our simulation running in real-time. On the left, the vehicle is shown in the environment, and on the right, the latest RGB image and depth image are shown on top, and the local and global costmap with planned paths overlaid are shown on the bottom.
Figure 11: Comparison of paths taken by baseline approach vs. our approach for two scenarios. Both scenarios share the same global region and start and goal locations, marked by a green circle and start respectively, and contain a high-cost region, highlighted in yellow. In the first scenario (top), this high-cost region is filled with medium sized rocks, and the second scenario (bottom) features small rocks. In both scenarios, our approach (paths shown in blue) successfully routes the rover around the high-cost region after it is observed, whereas the baseline approach (paths shown in red) paths the rover through the high-cost region, causing it to become stuck and/or accumulate high-cost.
 BaselineOurs (replan)
Total path cost39.7713.44
Number of collisions4.30
Total time289.6 s\mathrm{s}323.6 s\mathrm{s}
Total path distance335.7 m\mathrm{m}378.6 m\mathrm{m}
Table 1: Comparison of path metrics for baseline and our approach (averaged over three runs each). The baseline approach plans paths through the global high-cost region, resulting in shorter paths with higher cost and collisions, whereas our approach re-routes the global plan on the fly, resulting in slightly longer paths but with lower cost and no collisions.
More in this topic

Robot Motion Planning

Robot motion planning finds safe, feasible paths and trajectories for robots moving through constrained, changing environments.

See all 271 papers

Explore similar work

CardsList

  1. Safe Navigation using Neural Radiance Fields via Reachable Sets

    Apr 29, 2026Omanshu Thapliyal, Malarvizhi Sankaranarayanasamy, Ravigopal VennelakantiConstrained Motion PlanningSafe Robot Navigation

    Researcher, Hitachi America Ltd. Santa Clara, USA

  2. Learning Traversability for Long Horizon Off-Road Navigation

    Jul 26, 2026Kasi Viswanath, Jason M. Gregory, Shaunak Kolhe +1Off-Road Robot NavigationGeospatial Representation Learning

    J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77840, USA · DEVCOM Army Research laboratory, Adelphi, MD, USA

  3. Learning to Drive on Mars: Visual Multimodal Traversability Estimation for Off-World Navigation

    Sep 21, 2026Darren Chiu, Cole Wilson, Andrei Tumbar +2Off-Road Robot NavigationSpace Robotics

    University of Southern California, Los Angeles, CA 90089, USA. · Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109, USA. · Washington State University, Pullman, WA 99164, USA.