PACER: Preference-Conditioned All-Terrain Costmap Generation
Luisa Mao, Garrett Warnell, Peter Stone, Joydeep Biswas
AI summary
Problem
Current terrain cost assignment methods are restricted to predefined semantic classes or require costly retraining for new preferences, preventing rapid adaptation to novel terrain preferences during robot deployment.
Approach
PACER uses a neural network that processes a bird’s-eye view image alongside a small, user-provided preference context of terrain patch pairs to generate a corresponding costmap, trained via a staged pipeline on real and synthetic data.
Key results
- Rapid adaptation to novel user preferences at deployment time
- Superior generalization to unseen terrains over semantic and representation-learning baselines
- Generation of fine-grained, preference-aligned costmaps from single BEV images
- Open-sourced code and dataset for preference-conditioned costmap generation
Why it matters
Allows autonomous robots to dynamically align navigation paths with human operator preferences in diverse, unstructured environments without costly retraining.
Abstract
In autonomous robot navigation, terrain cost assign- ment 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 a staged training procedure leveraging real and synthetic data, we find that PACER is able to adapt to new user preferences at deployment time while also exhibiting better generaliza- tion to novel terrains compared to both semantics-based and representation-learning approaches. We release our code and dataset at https://github.com/ut-amrl/PACER RAL 2025.git