Forcesteed Robotics, AIST co-author paper on safety of generative AI-powered delivery robots

Forcesteed Robotics Co., Ltd. (Koto-ku, Tokyo) will present at The Robotics Society of Japan research conducted jointly with researchers from the National Institute of Advanced Industrial Science and Technology (AIST) on identifying and mitigating risks in indoor delivery robots equipped with physical AI that uses generative AI. The presentation is scheduled for 10:40 a.m. on Sept. 3, 2026, at the Natural Science Building at Kanazawa University. The paper is titled “Risk Identification and Risk Mitigation Measures for Indoor Delivery Robots Equipped with Physical AI Using Generative AI.”

Managing the uncertainty in robots’ decisions safely is essential to putting generative AI-equipped robots into practical use. The joint research presents specific verification methods and approaches to risk mitigation, and could affect the design and regulation of future physical AI-equipped robots.

The presentation is notable for testing the safety of “physical AI,” which uses generative AI to guide robot behavior, in a setting close to actual operations. Traditional robot safety has centered on deterministic controls, such as limiting operating areas and using sensors to avoid collisions. Generative AI, by contrast, has probabilistic variability in its perception and decisions, and errors can directly affect behavior in the physical world. The study set a family restaurant as a specific environment where people and robots share space, and tested not only whether the robot could move but also scenarios that could arise in an actual service setting, including responses to people and obstacles and the management of instruction authority. In this respect, the research seeks to move discussions of safety for generative AI-equipped robots from abstract risk analysis toward empirical data. In previous coverage, we reported that HAI ROBOTICS’ HaiPick Climb surpassed 10,000 cumulative orders and is advancing end-to-end automation with carton handling, and that Logic Robotics announced the ceiling-mounted robot “Octopus”, which transports goods without taking up aisle space. These developments show the expansion of logistics and delivery robot deployments and the emergence of diverse forms, while both involve the challenge of operating safely in the same spaces as people. The joint research by Forcesteed Robotics and AIST addresses that challenge from the perspective of generative AI, and the importance of such safety testing is expected to grow as delivery robots become more widely deployed in society. In terms of implications for the industry, the research is a collaboration between a public research institution, AIST, and a startup. If an assessment method for safety is established, it could become a common benchmark for robot manufacturers and system integrators. In particular, risk assessment when incorporating generative AI decisions into robot controls is currently handled separately by each company, and standardization has not advanced. Whether the research leads to guidelines or standards beyond its presentation at the society will be a key issue. Unresolved questions include the fact that individual test results and details of risk mitigation measures will be presented at the society, leaving specific risk cases and the effectiveness of countermeasures unclear at this point. Attention will also focus on the results of planned additional testing in areas including reducing recovery time after an anomaly, predicting human behavior, monitoring the robot’s own condition and making decisions based on the reliability of recognition results.

The research assumed meal service at a family restaurant and conducted tests in both a 3D simulation environment and a mock environment built at AIST. Scenarios included basic delivery from the kitchen to a table; responding to people in the robot’s path or crossing its path; providing service based on surrounding circumstances, such as a child waving; and managing authority by distinguishing instructions from employees and customers. The setup made it possible to check not only whether the robot moved but also how the generative AI perceived and judged its surroundings. It also allowed the researchers to isolate whether a failure to perform as expected stemmed from environmental perception, the generative AI’s decision or robot control. Testing confirmed cases in which the robot did not perform as expected because of differences between simulation and real environments, the robot’s size and range of motion, the surrounding environment, and combinations of AI perception and decision-making. Individual test results and details of risk mitigation measures will be reported at the society. The co-authors are Takaaki Namba (AIST), Yuta Iwase (seconded to AIST from Ad-Sol Nissin Corporation), Hiroki Tsukita (Forcesteed Robotics), Hiroyuki Osawa (Forcesteed Robotics) and Koichi Konishi (AIST). Forcesteed Robotics conducted verification experiments using the indoor delivery robot in this research.

The joint work by AIST and a Japanese startup to assess the safety of generative AI-equipped robots could contribute to the development of safety standards for Japan’s robotics industry. In particular, testing in settings where people and robots coexist, such as restaurants and commercial facilities, matches the characteristics of Japan’s service robot market and could provide a foundation for future domestic deployment.