Skip to main content
Back to timeline
arXivSource publication:

CityNav-VLA splits urban drone navigation into two frames, cutting endpoint error 21.2% on unseen tests

Synopsis

Language-goal urban drone navigation asks a UAV to reach a destination described in natural language, often outside its current view. CityNav-VLA separates a persistent north-up map memory from body-frame execution: a frozen multimodal backbone reads the instruction, current observations and map imagery, retrieves memory evidence, and predicts a target in map coordinates that a deterministic controller converts into body-frame commands. On the CityNav unseen test split it lowers endpoint error by 21.2% over the best prior result; for success rate and path efficiency on that dataset, the abstract and the introduction report different relative changes.

Source-provided article image: Allocentric Memory, Egocentric Acting: A Dual-Frame VLA for Urban Aerial Navigation
arXiv

Interpretation

Urban drone navigation stores spatial evidence on a north-up map whose position, orientation and scale stay fixed for the whole flight, predicts the target in map coordinates, and converts actions to the body frame using the current pose. Earlier language-goal aerial navigation either reasoned in first-person observations that change with viewpoint or fused map features implicitly inside a policy network, without explicit spatial reasoning over long-range targets. On the CityNav unseen test split, CityNav-VLA lowers endpoint error by 21.2% over the best prior result; on refined CityNav the endpoint-error reduction relative to HETT* is 6.2%. For success rate and path efficiency on the same dataset, the abstract gives 15.8% and 30.9% while the introduction gives 26.1% for path efficiency, so the two figures disagree.

Each step's query combines instruction context, UAV state and motion history, retrieves relevant evidence from map memory through cross-attention, and outputs a waypoint in map coordinates. Map memory places accumulated RGB observations, terrain height, landmark contours, explored area and flight trajectory on one grid, and metric coordinates anchor each cell to a fixed physical location. In a cumulative ablation on all 2,697 refined CityNav unseen-validation episodes with early goal commitment disabled, adding navigation-map features gives the largest endpoint-error reduction, and metric coordinates further improve endpoint accuracy while their success-rate and path-efficiency gains are not statistically significant.

Early goal commitment fixes the coordinate-wise median of the first predictions as the execution target at inference time, adding no trainable parameters. Without commitment the controller follows the current prediction at every step, so the target drifts as observations change over long horizons. On the CityNav unseen test split, enabling commitment on the same checkpoint raises success rate from 22.16% to 26.59% and path efficiency from 21.15 to 25.70, with 496 episodes becoming successful and 261 losing success; on unseen validation the success gain is only 0.21 percentage points and not significant.

Perspective

Evaluation covers the validation and unseen test splits of CityNav and refined CityNav, with endpoint error, success within 20 m, trajectory coverage and path efficiency as metrics; execution applies simultaneous forward, lateral and vertical displacements per step with heading held fixed over 120 decisions per episode.

On unseen validation, early goal commitment raises success rate by only 0.21 percentage points, not significantly, and lowers trajectory coverage by 0.71 points; commitment can also retain an incorrect early estimate, preventing recovery when later observations reveal a better destination, and the absence of explicit stopping lets the UAV pass near the target and then move away. For success rate and path efficiency on the same dataset, the abstract and the introduction report different relative changes, so any citation should name its source.