TARIC: Memory-Augmented Traversability-Aware Outdoor VLN under Interrupted Semantic Cues

Researchers present TARIC, a vision-language navigation framework that enables autonomous robots to complete outdoor navigation tasks despite interruptions in visual goal cues. The system combines semantic understanding with real-time traversability analysis to maintain feasible guidance during extended periods without visible landmarks, achieving 40% real-world success compared to 17.5% for existing methods.

TARIC addresses a fundamental challenge in autonomous robotics: maintaining goal-directed navigation when visual cues become unavailable, occluded, or move outside the robot's field of view. This problem is particularly acute in long-range outdoor environments where semantic landmarks cannot be guaranteed to remain visible throughout a journey. Previous memory-based approaches attempted to solve this by storing historical cues, but failed when remembered directions became physically infeasible due to terrain obstacles, causing robots to enter degraded states of backtracking and aimless wandering.

The innovation lies in grounding semantic understanding with real-world traversability constraints. Rather than treating traversability as a post-hoc safety filter, TARIC integrates it as a foundational component of navigation guidance. The system extracts bearing information from visible goal cues and translates these into executable headings using continuous near-field terrain analysis. During extended cue-free phases, the framework maintains a world-aligned 3D memory of previous observations with uncertainty quantification, allowing the robot to generate stable guidance that remains physically reachable as terrain conditions change.

For the robotics and autonomous systems industry, this represents a meaningful advance in real-world deployment viability. The 2.3x improvement in real-world success rates over baselines demonstrates practical impact beyond simulation metrics. The research validates that multi-modal integration—combining vision-language models with terrain understanding—outperforms semantic-only approaches in open-world conditions. This has implications for autonomous delivery, search-and-rescue, and exploration applications where visual landmarks cannot be guaranteed. Continued refinement of such systems could accelerate commercial deployment timelines for outdoor autonomous platforms.

• →TARIC framework enables robots to navigate outdoor environments by maintaining goal-directed guidance even when visual cues disappear or become obstructed.
• →The system achieves 40% real-world success rate versus 17.5% for existing methods by integrating terrain traversability analysis with semantic understanding.
• →Traversability constraints are treated as fundamental to navigation stability rather than merely local safety considerations.
• →A world-aligned 3D memory with uncertainty-aware readout prevents guidance degradation during forced detours caused by terrain obstacles.
• →Testing on both quadrupedal and wheeled platforms over 600-1000 meter routes demonstrates robustness across different robot morphologies.