acceptodds
Desk rejected at ICLR 2027

TRACE: Task Completion Determination for Aerial Navigation with Semantic Evidence

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied navigation can fail even when an agent reaches the goal region: execution may continue until the agent terminates elsewhere. We study task completion determination in aerial navigation, asking how target-related observations gathered during execution can inform an explicit termination decision. We introduce TRACE, **T**ermination via **R**epeated **A**ssessment of **C**ompletion **E**vidence, a training-free, plug-and-play semantic termination module that attaches to an existing navigation policy without changing its weights. A pretrained vision-language model (VLM) checks the current image against a target description at fixed action-step intervals and reports whether the target is visible. TRACE adds a STOP after consecutive positive reports, while preserving the navigator's own STOP decisions and otherwise leaving movement actions unchanged. The module uses no privileged goal-distance input. Visibility reports guide stopping, while final success is evaluated separately at the terminal position. We evaluate TRACE with frozen OpenFly-Agent and three pretrained VLMs serving as semantic judgers across two simulators. On routes starting outside the goal region, consecutive confirmation yields higher mean success than first-positive stopping across all matched comparisons, while dataset-specific selected configurations also outperform the original navigator with all three VLMs, improving successful termination within the goal region.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language Navigation, Aerial Navigation, Embodied AI, Temporal Evidence Accumulation, Reliable Stopping, Test-Time Inference

Related papers

Discussion (0)

Sign in to comment.