PACER: Preference-conditioned All-terrain Costmap Generation

TL;DR

PACER addresses the challenge of rapidly adapting terrain costs to operator preferences by learning a preference-conditioned BEV costmap from a deployment-time preference context. It casts costmap generation as learning a function $R:\mathcal{I}\times\mathcal{H}\to\mathcal{C}$ that aligns with the unknown human cost $H:\mathcal{T}\to\mathbb{R}$ under equivalence and partial-ordering constraints, enabling navigation in $SE(2)$. The authors propose a two-encoder encoder–decoder network that processes a BEV image and a structured preference context $\hat{H}$, trained with pixel-wise loss across real, shuffled, and synthetic data in three phases. Empirical results from aerial-map simulations and real-robot deployments show that PACER generalizes to unseen terrains, adapts to new preferences without fine-tuning, and outperforms semantics-based and representation-learning baselines in producing navigation-friendly costmaps.

Abstract

In autonomous robot navigation, terrain cost assignment is typically performed using a semantics-based paradigm in which terrain is first labeled using a pre-trained semantic classifier and costs are then assigned according to a user-defined mapping between label and cost. While this approach is rapidly adaptable to changing user preferences, only preferences over the types of terrain that are already known by the semantic classifier can be expressed. In this paper, we hypothesize that a machine-learning-based alternative to the semantics-based paradigm above will allow for rapid cost assignment adaptation to preferences expressed over new terrains at deployment time without the need for additional training. To investigate this hypothesis, we introduce and study PACER, a novel approach to costmap generation that accepts as input a single birds-eye view (BEV) image of the surrounding area along with a user-specified preference context and generates a corresponding BEV costmap that aligns with the preference context. Using both real and synthetic data along with a combination of proposed training tasks, we find that PACER is able to adapt quickly to new user preferences while also exhibiting better generalization to novel terrains compared to both semantics-based and representation-learning approaches.

Figures (6)

• Figure 1: Given an input image $I$ and a preference context $\hat{H{}}$ of $n$ ordered pairs of terrain patches where the left terrain is more preferred than the right, pacer generates a costmap consistent with this preference. Changing the preference context leads to changed terrain costs, which results in a different plan aligned to the new operator preference. The paths planned according to the different preferences are shown above. In the costmap, black represents low cost and white represents high cost.
• Figure 2: Relationships between spaces of Terrains, Image Observations, and Costmaps. There exists a hidden "true" costing function based on human preferences directly on terrains. pacer approximates this function from visual observations of terrains.
• Figure 3: Overview of the dataset structure. Each training example contains a preference context, image, and target costmap. We vary the preferences and images, resulting in a large combinatorial dataset despite the relatively small amount of real recorded data. In a later training phase, we also augment with synthetic data by artificially finding and replacing certain terrain types with synthetic terrain textures. The real-valued costs assigned to terrain types based on an input total ordering are shown in the Generate Examples procedure, where black represents low cost and white is high cost.
• Figure 4: The effect of changing preference on path planning is shown on the left image of an aerial map. The blue path corresponds to the middle costmap and the red path corresponds to the right costmap. The generated costmaps reflect the preferences provided in the context. Dark purple corresponds to low cost and yellow is high cost.
• Figure 5: Examples of several paths planned using our method in an urban unseen environment and their corresponding preference contexts. Here, we visualize the large scale of these simulated deployments, and the diversity of visual terrain appearances.

Theorems & Definitions (2)

• Theorem 1
• proof