Improving Reliable Navigation under Uncertainty via Predictions Informed by Non-Local Information

TL;DR

This work addresses reliable long-horizon navigation under uncertainty in partially mapped environments by enabling predictions about unseen space that incorporate non-local information. It extends the Learning over Subgoals (LSP) planning framework with a Graph Neural Network (GNN) backbone, forming LSP-GNN, which operates on a skeleton-based graph of the environment to predict subgoal properties $P_S$, $R_S$, and $R_E$ and thus informs high-level planning. The method preserves reliability guarantees while achieving significant performance gains across three challenging environments, including large-scale university floorplans, with improvements up to 14.9% over local-information baselines. This approach demonstrates that maintaining non-local knowledge via structured graph representations can markedly enhance building-scale navigation under uncertainty, with potential for richer sensory integration in future work.

Abstract

We improve reliable, long-horizon, goal-directed navigation in partially-mapped environments by using non-locally available information to predict the goodness of temporally-extended actions that enter unseen space. Making predictions about where to navigate in general requires non-local information: any observations the robot has seen so far may provide information about the goodness of a particular direction of travel. Building on recent work in learning-augmented model-based planning under uncertainty, we present an approach that can both rely on non-local information to make predictions (via a graph neural network) and is reliable by design: it will always reach its goal, even when learning does not provide accurate predictions. We conduct experiments in three simulated environments in which non-local information is needed to perform well. In our large scale university building environment, generated from real-world floorplans to the scale, we demonstrate a 9.3\% reduction in cost-to-go compared to a non-learned baseline and a 14.9\% reduction compared to a learning-informed planner that can only use local information to inform its predictions.

Figures (12)

• Figure 1: Overview: non-local information is often essential for good navigation in a partial map. Our LSP-GNN approach uses a graph neural network to make predictions about unseen space via both local and non-local information and integrates these into the Learning over Subgoals model-based planning abstraction pmlr-v87-stein18abradley2021 to improve reliable navigation.
• Figure 2: Our robot's actions correspond to boundaries between free and unseen space. The robot can leave observed space through either boundary: via subgoal $s_1$ or $s_2$. Upon selecting action $a_2$, the robot reaches the goal with probability $P_S$ and incurs an expected cost $R_S$, or is turned back (probability $1-P_S$), accumulates cost $R_E$ and selects another action.
• Figure 3: Low cost navigation in our J-Intersection environment requires non-local information. When the goal is either on left or right from the intersection, we need the non-local information from the start position to decide correctly at the intersection. Choosing always left or right or even choosing one color over another will not reliably succeed.
• Figure 4: Graph representations of the environment for our graph neural net are computed from the partial map. We use an image skeleton zhang1984 to generate a graph from the partial occupancy grid. See Sec. \ref{['sec:generate-graph']} for details.
• Figure 5: Planned trajectories of the bench-marked planner approaches. J-intersection environment where the goal is on the right. The left column shows the optimal trajectory (planned using the underlying known map). The middle column shows the same trajectory of both the non-learned baseline and LSP-Local where they make a systematic choice. The right column shows the trajectory planned by LSP-GNN that is similar to the optimal one.