Fast Navigation Through Occluded Spaces via Language-Conditioned Map Prediction
By: Rahul Moorthy Mahesh , Oguzhan Goktug Poyrazoglu , Yukang Cao and more
In cluttered environments, motion planners often face a trade-off between safety and speed due to uncertainty caused by occlusions and limited sensor range. In this work, we investigate whether co-pilot instructions can help robots plan more decisively while remaining safe. We introduce PaceForecaster, as an approach that incorporates such co-pilot instructions into local planners. PaceForecaster takes the robot's local sensor footprint (Level-1) and the provided co-pilot instructions as input and predicts (i) a forecasted map with all regions visible from Level-1 (Level-2) and (ii) an instruction-conditioned subgoal within Level-2. The subgoal provides the planner with explicit guidance to exploit the forecasted environment in a goal-directed manner. We integrate PaceForecaster with a Log-MPPI controller and demonstrate that using language-conditioned forecasts and goals improves navigation performance by 36% over a local-map-only baseline while in polygonal environments.
Similar Papers
NavForesee: A Unified Vision-Language World Model for Hierarchical Planning and Dual-Horizon Navigation Prediction
Robotics
Helps robots plan and imagine to navigate better.
CogniPlan: Uncertainty-Guided Path Planning with Conditional Generative Layout Prediction
Robotics
Helps robots explore and move in new places.
Language Conditioning Improves Accuracy of Aircraft Goal Prediction in Untowered Airspace
Robotics
Helps planes understand where others are going.