Incident Overview
The incident occurred when all three of the robot’s sensor systems—likely including LiDAR, cameras, and ultrasonic sensors—failed to detect the transparent glass structure. Serve Robotics confirmed the failure and later disclosed that the robot’s perception software and hardware had been updated to mitigate risks associated with transparent obstacles. The company also adjusted operational protocols to restrict robot movement near bus shelters.
Two days later, on March 24, 2026, a Coco Robotics delivery robot experienced a similar incident in Chicago’s Old Town neighborhood. The robot struck a glass bus shelter around 4 p.m., causing comparable damage to the shelter’s panels and spreading glass fragments across the sidewalk. Unlike the Serve Robotics incident, Coco Robotics did not specify sensor failures but described the crash as a rare occurrence. The company took responsibility for cleanup and repairs, stating no injuries were reported. Both incidents were confirmed by the organizations involved and documented by the OECD AI Incidents and Hazards Monitor (AIM).
The two incidents share a common root cause: the inability of autonomous delivery robots to reliably detect transparent obstacles, such as glass bus shelters, using their onboard perception systems. This failure highlights a critical limitation in current robotic perception technologies, particularly in urban environments where glass structures are prevalent.
The damage in both cases was limited to property—specifically, the destruction of glass panels and the need for sidewalk cleanup. Chicago officials reviewed permits for autonomous delivery robots following the incidents, though no immediate policy changes were announced. Both Serve Robotics and Coco Robotics issued public statements acknowledging responsibility and committing to further software and operational improvements.
The failures underscore the need for improved object detection algorithms, particularly for non-reflective or transparent materials, which are not yet adequately addressed in commercial robotic navigation solutions.
Possible Root Cause
The primary root cause of both incidents was the failure of the robots’ perception systems to detect transparent glass obstacles. However, glass surfaces present a unique challenge because they do not reflect LiDAR beams effectively and may appear as empty space in sensor data. Cameras, while capable of detecting edges, struggle with transparent materials due to their lack of distinct visual texture or contrast.
Serve Robotics explicitly stated that all three of its sensor systems failed to detect the bus shelter’s glass wall during the March 22 incident. This suggests a systemic limitation in sensor fusion—the process of combining data from multiple sensors to create a coherent understanding of the environment. When transparent objects like glass are involved, the fusion algorithm may not generate an accurate representation of the obstacle, leading to navigation errors. The company’s subsequent software updates indicate an acknowledgment of this flaw and an attempt to address it through improved perception algorithms.
Coco Robotics did not provide a detailed technical explanation for its robot’s failure, but the incident followed a nearly identical pattern to Serve Robotics’ crash. This repetition strongly implies a shared vulnerability in current robotic perception systems when encountering transparent obstacles. Both companies’ responses suggest that their initial sensor configurations were insufficient for reliably identifying glass surfaces.
The incidents also point to broader challenges in autonomous robotics, including the reliance on probabilistic sensing and the difficulty of training machine learning models to recognize edge cases like transparent materials. Glass, by contrast, lacks these properties, forcing robots to rely on indirect detection methods—such as identifying the frame or edges of the structure—which may not always be present or detectable in real-world conditions.
Takeaways
The Chicago bus shelter incidents reveal a critical limitation in the current state of autonomous delivery robot perception systems: the inability to consistently detect transparent obstacles. This failure is not unique to Serve Robotics or Coco Robotics but reflects a broader industry challenge that could affect other autonomous mobile robots operating in urban environments. The incidents serve as a case study in the risks associated with deploying AI-driven systems in complex, real-world settings where sensor limitations may lead to unintended consequences.
For robotics developers, the incidents underscore the need for more robust perception algorithms that can handle edge cases, such as transparent or semi-transparent materials. Current sensor fusion techniques may require enhancement to improve detection accuracy for such obstacles.
Chicago officials’ review of permits suggests that local authorities may begin to scrutinize the safety protocols of autonomous delivery services more closely. Companies deploying such robots should proactively assess environmental risks and implement conservative operational boundaries—such as avoiding high-risk areas or slowing down near potential obstacles—until perception systems are fully validated.
Both Serve Robotics and Coco Robotics acknowledged the failures and committed to improvements, but the lack of detailed technical explanations in Coco’s case left some questions unanswered.
Finally, the repeated nature of these incidents suggests that the robotics industry may need to adopt a more cautious approach to urban deployments until perception systems can reliably handle all types of obstacles. While autonomous delivery robots offer significant efficiency benefits, their limitations—particularly in edge-case scenarios—must be addressed before widespread adoption can be considered safe.
The property damage caused by these incidents, while not life-threatening, demonstrates the potential for escalating harm if similar failures occur in more dangerous contexts. For example, a robot failing to detect a transparent barrier in a pedestrian-heavy area could lead to injuries. The takeaway for stakeholders is clear: perception system reliability must be prioritized, and operational safeguards must be implemented to prevent avoidable collisions until technology matures.
Sources
- incidentdatabase.aihttps://incidentdatabase.ai/cite/1567/
- incidentdatabase.aihttps://incidentdatabase.ai/cite/1568/
Incident summary from the AI Incident Database, used under CC BY-SA 4.0.