RobotWorld

Why Delivery Robots Still Struggle With Sidewalks and Streets

By RobotWorld·9/16/2026

Autonomous delivery robots have a promise that's hard to argue with: zero-emission, last-mile logistics that frees up road space and cuts delivery costs. They've been spotted on college campuses, in suburban neighborhoods, and outside grocery stores across the US, Europe, and Asia. Yet a growing body of user complaints, operator reports, and independent observations points to a stubborn reality — these machines frequently don't behave the way people expect them to in shared public spaces.

The core issues aren't exotic or mysterious. They're the exact friction points you'd predict once you put a wheeled robot on a sidewalk full of pedestrians, bikes, uneven pavement, and ambiguous social cues.

The Rules of the Road (and Sidewalk) Are Harder Than They Look

Road rules for cars are codified and largely consistent. Sidewalk norms, on the other hand, are a combination of local ordinance, unwritten social custom, and real-time negotiation between strangers. Humans navigate this instinctively — making eye contact, reading body language, stepping aside, or signaling intent with a head nod. Robots can't do any of that yet with reliability.

Current delivery robots rely on a combination of cameras, ultrasonic sensors, and LiDAR to detect obstacles and plot paths around them. When an obstacle is static — a parked car, a mailbox, a wall — they generally cope well. The trouble comes with dynamic, socially embedded situations: a group of people standing in conversation, a cyclist who makes an unexpected turn, a dog on a long leash, or a construction zone where half the sidewalk is blocked with no formal detour. These scenarios require prediction and social inference, not just detection.

The result is robots that stop abruptly, take confusing detours, or — in the cases that generate the most complaints — simply block foot traffic while they wait for a situation to resolve. None of this is safe, and none of it builds public trust.

The Sensor and Compute Gap

Part of the issue is hardware capability at the price points these services need to operate commercially. High-fidelity perception — the kind used in research-grade autonomous vehicles — is expensive. Delivery robots are cost-constrained machines, so their sensor suites are often less sophisticated than what a premium drone or inspection robot would carry.

This matters because robust real-world navigation depends on rich, accurate environmental data fused from multiple sources. Platforms designed for demanding autonomous operation — like the Unitree B2 quadruped, built for rough terrain and equipped for autonomous navigation — illustrate the level of engineering required for a machine to handle truly unstructured environments. Translating that capability into a wheeled delivery bot at a commercially viable cost remains an open engineering challenge.

On the compute side, running real-time perception, path planning, and behavioral inference on an embedded system is genuinely difficult. Edge AI platforms like the NVIDIA Jetson Orin Nano Super are designed precisely to address this bottleneck — delivering substantial AI inference capability in a compact, low-power form factor suitable for mobile robots. But integrating that compute power with the right sensor suite, trained on sufficiently diverse real-world data, is where many commercial deployments fall short.

The Training Data Problem

Machine learning models are only as good as what they've been trained on. If a delivery robot's navigation AI has learned primarily from controlled test environments or a limited set of neighborhoods, it will struggle when it encounters the full messiness of real public space — unfamiliar curb cuts, rain-slicked pavement, construction diverting pedestrian flow, or crowds behaving in culturally specific ways.

This is a data diversity problem. It's why robots that perform well in a sunny California suburb may behave erratically in a snowy Midwestern city, or why a bot tested in a low-pedestrian-density environment can become confused on a busy urban block. The solution — gathering vastly more varied training data, and continuously retraining models — is straightforward in principle but slow and expensive in practice.

The Regulatory Patchwork

Technical challenges aside, delivery robots also operate in a fragmented regulatory environment. Rules vary not just by country but by city and sometimes by neighborhood. Speed limits for sidewalk robots, weight restrictions, areas where autonomous operation is permitted, and requirements for remote human oversight all differ significantly across jurisdictions. Operators building fleets at scale have to design for the most restrictive environment they serve while also complying with every local variation — a compliance burden that slows both deployment and iteration.

What Good Looks Like

None of these problems are unsolvable, and progress is real. Researchers are developing better social navigation models — systems that allow robots to predict and adapt to human movement rather than just react to it. Sensor fusion techniques are improving, combining camera, LiDAR, and radar data more effectively. And as edge AI hardware matures, the compute available to an affordable mobile robot is increasing substantially.

Indoor environments show what becomes possible when the problem is constrained. The Pudu BellaBot, a commercially deployed food delivery robot used in hospitality settings, navigates restaurants and hotel corridors reliably using dual-SLAM (LiDAR and visual) positioning — because the environment is controlled, mapped in advance, and largely free of unpredictable pedestrian behavior. The outdoor sidewalk problem is a fundamentally harder version of the same challenge.

The Trust Equation

Technology aside, there's a human dimension here that operators underestimate. Public acceptance of delivery robots depends heavily on whether people feel those machines behave sensibly and respectfully in shared spaces. A robot that blocks a wheelchair user's path, startles a child, or confuses an elderly pedestrian doesn't just fail a delivery — it erodes the broader social license that allows these fleets to operate at all.

Getting delivery robots right isn't purely a hardware or software question. It's a question of designing machines that can participate in public life as predictable, considerate, and safe members of a shared environment. That bar is higher than it might appear — and the industry is still working its way toward it.

References

This article was drafted with AI assistance and reviewed before publishing.

Related reading