New AI Model Teaches Delivery Robots to Navigate City Streets Autonomously

Researchers have developed a machine-learning system that enables small delivery robots to navigate crowded urban streets by understanding visual scenes and following natural language route instructions simultaneously. The model, called UrbanVLA, bridges a major gap in autonomous robot development by handling both the fine details of movement (dodging obstacles, reaching specific spots) and the broader challenge of understanding where to go across large city areas.

The dual challenge of city navigation#

Existing navigation systems typically specialize in either short-range tasks—like avoiding a pedestrian or reaching a nearby goal—or long-range route planning in controlled environments. Real-world urban delivery requires robots to excel at both. A robot must simultaneously understand visual cues from its camera, align itself with a mapped route, and execute safe movements through unpredictable human-filled spaces. This combination remains rare in deployed systems.

How UrbanVLA works#

The model operates by matching visual observations from the robot's camera against noisy, imprecise waypoint data describing the intended route. Once aligned, it calculates a safe trajectory for the robot to follow. Rather than relying on a single training method, the researchers used a two-phase approach: first, they used supervised learning on simulated environments and trajectories extracted from online videos to teach the model the fundamentals. Then they applied reinforcement learning on a blend of simulation and real-world footage, letting the system refine its strategies through trial and feedback in realistic conditions.

Real-world results#

Testing on the SocialNav benchmark—which evaluates robots navigating around people—UrbanVLA outperformed competing systems by more than 55%. In actual urban deployments, the model demonstrated both scalability across large-scale city environments and resilience when facing unexpected real-world variations. This dual success suggests the approach may be ready for practical deployment in delivery and other micromobility applications.