Delivery Robots Struggle With Edge-Case Safety
The Chicago Sun-Times reports two delivery robots struck CTA bus shelters in Chicago in March, raising safety concerns. According to the Sun-Times, a Serve Robotics unit went through the glass of a West Town bus shelter on March 22 and a Coco Robotics unit hit a shelter near North Larrabee Street on March 24. The newspaper cites a 2025 University of Pennsylvania engineering study recommending improvements to internal sensors and attention to an optical illusion that makes clean glass harder for robots to detect. The Sun-Times also quotes University of Pennsylvania dean Vijay Kumar describing delivery robots as encountering "edge cases," and a local resident expressing concern that people could have been hurt.
What happened
The Chicago Sun-Times reports that on March 22 a Serve Robotics delivery robot went through the glass of a bus shelter in West Town, and on March 24 a Coco Robotics robot collided with a bus shelter near North Larrabee Street in Lincoln Park. The Sun-Times says the incidents attracted viral attention and prompted public concern, including a petition to pause the city trial program begun in 2022. The article cites a 2025 study by engineering experts at the University of Pennsylvania that the newspaper frames as offering technical suggestions for reducing such failures.
Technical details
The Sun-Times reports that Coco Robotics uses 360-degree cameras, laser sensors and artificial intelligence, while Serve Robotics uses AI plus multiple sensor types and both companies employ remote human operators to monitor and intervene. The newspaper quotes Ali Kashani, CEO of Serve Robotics, saying three internal sensors failed simultaneously and that video from the Serve robot's viewpoint showed the bus shelter glass was difficult to see. The Sun-Times also notes one company ran an advertisement in which the robot "apologize[d]" after the incident.
Industry context
Editorial analysis: Industry observers and researchers commonly describe perception failures as "edge cases," where robots encounter physical situations not well represented in training data or simulators. Clean, transparent surfaces produce low-contrast reflections and refractions that can defeat vision-based detectors, and simultaneous sensor faults or degraded sensor fusion increase the chance of unhandled failure modes. Remote teleoperation mitigates some risks but introduces latency and situational awareness limits.
What to watch
Editorial analysis: Practitioners and municipal observers will likely track post-incident investigations, any data released about sensor logs or failure modes, updates to trial permitting or public-hearing schedules, and whether independent audits of perception and safety systems are requested. Broader indicators include incidence rates of collisions, third-party validation of sensor-fusion stacks, and published mitigation strategies for transparent-surface detection.
Scoring Rationale
The incidents are materially relevant to practitioners working on perception, sensor fusion, and urban robotics safety, but they do not represent a sector-wide technical breakthrough. The story flags operational risks and testing gaps that affect deployment timelines and safety evaluations.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
