Context-Aware Navigation
Leveraging context for robotic navigation in unknown environments
This project develops a decentralized framework for context-aware navigation, enabling embodied agents to perform complex path-finding and search tasks in unknown environments without prior maps. By leveraging Graph Attention Networks to encode environmental context, the framework allows robots to reason about the global structure of a space from local observations. This enables a suite of navigation capabilities—from zero-shot exploration to adaptive prior-based path-finding—that outperform traditional geometric planners in both computational efficiency and success rate.
Context LearningTest-Time AdaptationMultimodal FusionPath-Finding Priors
Contributions
What this work contributed, drawn from the papers below.
-
01
Mapless navigation from an online belief
The robot navigates to a predefined destination using partial knowledge of the environment that it updates online, rather than relying on a prior map.
Context-Aware DRL, CoRL 2023 -
02
A context formed over the entire known area
The policy network forms a context of the robot's belief over everything it has seen so far, using it to reason about long-term efficiency while sequencing short-term movements.
Context-Aware DRL, CoRL 2023 -
03
Graph rarefaction for large-scale planning
A graph rarefaction algorithm thins the belief graph so decision-making stays efficient as environments grow.
Context-Aware DRL, CoRL 2023 -
04
Measured against classical planners
Across hundreds of test scenarios the approach cut average travel time by up to 61.4% and average planning time by up to 88.2% against D* Lite and BIT*, and was validated in high-fidelity Gazebo simulation and on hardware without further training or tuning.
Context-Aware DRL, CoRL 2023 -
05
Satellite imagery as a search prior
For outdoor visual search, satellite imagery can inform high-level strategy even when its resolution is too low to recognise the target. Existing planners either assume no prior, or use priors without accounting for how they were produced.
Search-TTA, CoRL 2025 -
06
A satellite encoder aligned to CLIP
A satellite image encoder is pretrained to align with CLIP's visual encoder, so it outputs probability distributions of target presence that a planner can search over.
Search-TTA, CoRL 2025 -
07
Test-time adaptation against VLM hallucination
Search-TTA refines CLIP's predictions during the search itself using uncertainty-weighted gradient updates inspired by Spatial Poisson Point Processes, correcting priors that start out wrong. The plug-and-play interface accepts image, text, or sound and works with planners including reinforcement-learning ones.
Search-TTA, CoRL 2025 -
08
AVS-Bench, and results on real hardware
AVS-Bench, a visual search dataset of 380k images and taxonomy data drawn from internet-scale ecological data, was curated to train and evaluate the method. Search-TTA improves planner performance by up to 30.0%, most of all when initial CLIP predictions are poor, performs comparably to significantly larger VLMs, generalises zero-shot to unseen modalities, and was deployed on a real UAV through hardware-in-the-loop testing.
Search-TTA, CoRL 2025