Terrain Costmap Generation via Scaled Preference Conditioning

TL;DR

SPACER addresses the dual challenge of generalizing terrain costmaps to unseen environments while enabling rapid test-time control of relative terrain costs. It achieves this by conditioning costmap generation on a scaled preference context and training on a synthetic dataset, using a latent UNet architecture and a pretrained Stable Diffusion VAE, with an inverse Bradley–Terry formulation to link preference strength to cost differences. The approach yields superior planning performance and costmap accuracy across diverse aerial environments, outperforming semantic segmentation and prior PbIRL baselines in most settings, and enabling fine-grained operator control without retraining. This has practical impact for autonomous off-road navigation, as operators can rapidly adapt to mission goals by adjusting preference strengths, with robust generalization to novel terrains.

Abstract

Successful autonomous robot navigation in off-road domains requires the ability to generate high-quality terrain costmaps that are able to both generalize well over a wide variety of terrains and rapidly adapt relative costs at test time to meet mission-specific needs. Existing approaches for costmap generation allow for either rapid test-time adaptation of relative costs (e.g., semantic segmentation methods) or generalization to new terrain types (e.g., representation learning methods), but not both. In this work, we present scaled preference conditioned all-terrain costmap generation (SPACER), a novel approach for generating terrain costmaps that leverages synthetic data during training in order to generalize well to new terrains, and allows for rapid test-time adaptation of relative costs by conditioning on a user-specified scaled preference context. Using large-scale aerial maps, we provide empirical evidence that SPACER outperforms other approaches at generating costmaps for terrain navigation, with the lowest measured regret across varied preferences in five of seven environments for global path planning.

Figures (7)

• Figure 1: Given an input image and preference over terrains, the model outputs a costmap. A scaled preference context is a set of pair comparisons of terrains (left patch preferred over right), scaled by a strength of preference$\alpha$. Varying the strength of a preference leads to different gradation of costs and resulting planned paths.
• Figure 2: Model architecture. Inputs are BEV image and variable-length preference context consisting of patch pairs and strengths. The conditioning latents are the concatentation of patch embeddings and strength embedding (calculated as a linear interpolation of strong and weak embeddings). The output is a costmap. updated
• Figure 3: The construction of the Target Costmap and Scaled Preference Context in the Training Example parallels the loss function. The loss function has two parts: preference reconstruction loss $\mathcal{L}_1$ and pixelwise reconstruction $\mathcal{L}_2$ against the target costmap. In $\mathcal{L}_1$, the preference context is constructed from the predicted costmap using the ground truth segmentation masks, and the predicted strengths are matched to the input (target) strengths. The $\mathcal{L}_2$ term keeps costs from exploding while the $\mathcal{L}_1$ term is critical for the model's preference-alignment ability. Note: $\mathcal{L}_1$ is shown in an equivalent form here for clarity, but we implement it as in eq. \ref{['eq:loss1']}.
• Figure 4: Examples of misclassified segmentation masks (left) and missed regions in segmentation masks (right). Each triplet consists of a spacer costmap, RELLIS-3D image, and RELLIS-3D segmentation mask. We find that the RELLIS-3D segmentation masks are not accurate and fine enough for a quantitative evaluation across the whole dataset.
• Figure 5: Qualitative comparison of costmaps generated by all methods. As preference strength varies across rows, pacer cannot adapt. As terrain visual appearances become more misaligned with their classes across columns (i.e. vegetation is not green), classification baselines cannot adapt. spacer produces costmaps and planned paths which are most aligned with the userspreferences