Autonomous AI systems test governance in physical environments

Autonomous AI systems begin to penetrate beyond software environments in warehouses, supply networks and public spaces. The development draws attention to the question whether the current AI rules cover systems operated in physical environments. Most existing AI governance frameworks focus on online damages and model results, including bias, misinformation and harmful content. Identified AI systems are risks in physical environments where failures can affect infrastructure, property or human security. Singapore's Infocomm Media Development Authority on 20. May Version 1.5 of your model AI Governance Framework for Agentic AI published. The Framework lays down guidelines for organisations that use AI agencies that plan, make decisions and take measures to achieve custom goals in several steps. The framework states that agents can interact with tools, external systems and other agents, including systems that update databases, write files, control devices or execute transactions. It lists access controls, monitoring and human approval as governance measures for deployment. AI injects into physical systems

At an AI summit in Singapore last week, the discussions on robotics and embodied AI focused on safety issues that are more often associated with aviation, industrial systems and oversight of critical infrastructures than with conventional software regulation. The speakers also discussed whether autonomous systems can function safely and reliably over longer periods of time in unpredictable real environments. Dr. Ya-Qin Zhang, founding Dean of the Institute of AI Industrial Research at Tsinghua University, said that embodied AI systems are strengthening the risks already associated with autonomous software. He said that failures could directly affect transport systems, drones, logistics networks and critical infrastructures. “Every risk in the digital field is enhanced in the physical area, and the physical area will have a physical consequence,” Zhang said to MLex at the edge of the summit. He added that vehicles, drones, intelligent networks and other infrastructures could come to light when AI systems are deeper embedded in physical processes. The speakers discussed reliability, operational monitoring and security following the provision as governance concerns. In the summit discussions, it has been pointed out to mission-based governance models based on simulation, telemetry and iterative tests and not just based on a unique certification. The IMDA framework also recommends step-by-step rollouts, continuous monitoring and further tests after deployment. It means that agents can interact dynamically with their environment and not anticipate all risks before publication. Monitoring will lead to a supply problem

Grab, which tests autonomous vehicles and delivery robots in the Punggol District in Singapore, said that the control of deployment depends heavily on simulation, testing and continuous monitoring. “We carry out many simulations, we carry out many tests in closed and open courses to ensure that our robots are reliable,” said Suthen Thomas Paradatheth, Chief Technology Officer of Grab, during one of the summits. “Before we scale to hundreds of robots, we make sure we crack it first in the simulation and with some robots,” he added. Grab also referred to monitoring systems that track robot performance and identify unexpected failures after use. “It could be a whole series of problems,” said Paradatheth. The IMDA Framework states that companies should evaluate agent AI applications based on data access, external system access, autonomy and task complexity. It also points to the extent and reversibility of agent actions, the participation of third parties and the overall complexity of the system. It is also recommended to limit agent access to tools and systems, to apply permissions with the lowest authorization and to define standard workflow instructions. Organizations should also set up mechanisms to connect agents offline for malfunctions. The responsibility extends to more players

MLex reported that several parties may be involved in the development, production and provision of embodied AI systems. These include AI developers, robotics manufacturers, semiconductor suppliers and infrastructure managers. MLex also noted that it can be more difficult to assign responsibility when systems are further adapted after deployment by software updates, telemetry and operating data. According to IMDA, organizations and people remain responsible for the agents' actions even if the agents act autonomously. The framework calls for clear responsibility in the entire Agent-KI value chain, from model and platform providers to suppliers, tool providers and end users. Applied Materials said that the large-scale use of robotics is also associated with semiconductor economy and system integration. Om Nalamasu, Chief Technology Officer of the company, said that robot systems will depend on better sensors, energy efficiency, advanced packaging and computer architecture. Nalamasu said that robotics systems would require specially designed designs adapted to certain industrial ecosystems, and not a single solution for all environments. Zhao Yuli, Chief Strategy Officer of Chinese Robotics Startup Galbot, said that Beijing prioritises the scope and industrial commercialisation through state-funded test environments, industrial partnerships and long-term financing initiatives. Galbot has used humanoid robot systems in retail, warehouse and pharmaceutical companies in China. These include autonomous shops that operate around the clock. Zhao said that semi-structured industrial environments are likely to become an early way to commercialization, as they provide better controllable operating conditions. Japan attaches more importance to defining standards, robotics records and security governance. Professor Yutaka Matsuo from the Graduate School of Engineering at the University of Tokyo referred to an “AI Association” project aimed at collecting 100,000 hours of robotic data to support robot foundation models. Matsuo also referred to the Japanese AI Safety Institute and the Hiroshima AI Process as part of more comprehensive efforts to develop governance standards for embodied AI systems with Singapore and other Asian countries. Singapore introduces agent controls

The Singapore framework sets four governance areas for Agent-KI. These include risk assessment in advance, human responsibility, technical controls and the responsibility of the end user. The framework describes it as an iterative process and not as a one-time assessment. The framework states that the human supervision for agent systems must be adapted, as a continuous review of all workflows on a large scale becomes impractical. It recommends human approval at important control points, including high-risk measures, irreversible measures and unusual behaviour. IMDA also identifies automation distortions and alarm fatigue as risks when people oversee capable agents. It is recommended to check the monitoring by indicators such as human overcontrol rates and reaction times and to use automated real-time monitoring to identify unexpected behavior. The framework states that users should be informed which actions an agent can perform, which data it can access and which responsibilities remain with the user. In addition, a training of employees for interaction between humans and agents, for supervision and professional skills required to evaluate the agent results is recommended. Companies test AI in regulated workflows

JPMorgan implements AI tools in its entire global investment banking business, said Paul Uren, Head of Investment Banking of the Bank for Asia-Pacific Space, to Reuters. The bank said that the tools help bankers access more information and synthesize it with internal systems. They are also used to prepare content and support customer loyalty. Jamie Dimon, CEO of JPMorgan, told Bloomberg News that the bank will hire more AI specialists and less traditional bankers. Reuters reported that global banks increase their AI investments, redesign the workforce and change the tasks. The Bank is also one of the selected organisations, Anthropic was allowed to use their myth cybersecurity model as part of a controlled initiative called Project Glasswing. According to Anthropic, Mythos can detect old vulnerabilities in browsers, infrastructure and software. Reuters reported that Goldman Sachs, Citigroup, Bank of America and Morgan Stanley also have access to Mythos or test it. The IMDA Framework includes a case study by OCBC Bank of Singapore to analyze the asset. The system analyzes income-related documents and creates a memo for the asset source. Credit, onboarding or risk decisions are not taken autonomously. In this case, the workflow is limited to autonomy at task level and is only executed when triggered by predefined workflows. At critical decision-making points, a human review is required and the final validation is for the designated examiners. Robots Hold Input into Industrial Use

According to one of Nikkei Research from 1 to 15 Reuters survey conducted in Japan in May, one third of the companies are already using or considering AI-based robots. As part of the survey, 492 companies were contacted, of which 220 responded under the condition of anonymity. About 4% of respondents stated that they already use AI robots, 5% plan their use and 25% consider this. The remaining 66% stated that there were no such plans. Manufacturers of transport equipment were the most active group in the survey: 80% already used AI robots or consider the use. In comparison, 94% of respondents in the wholesale trade indicated that they had no plans for using AI robots. 71% of the companies that use AI robots, plan their use or think about it chose production as an application. Another 19 % opted for dangerous tasks, while 11% chose customer-oriented services. The Japanese government expects AI robots to help combat the country's chronic labour shortage and strengthen its position in industrial robotics. Japan is home to robotics companies such as Fanuc, Yaskawa Electric and Kawasaki Heavy Industries, but is facing competition from China and the United States in AI-based robotics. Retail agents expand beyond search

Walmart has outlined plans for the use of Agent-KI in purchasing, employee, supplier and developer workflows. In July 2025, retailers announced plans for four AI-based ‘superagents’. They are directed to buyers, processors, suppliers and sellers as well as software developers. Walmart said these agents would become the main entry point for AI interactions between these groups. One of the tools, Sparky, is already available in Walmart's app as a generic AI-based shopping assistant. Hari Vasudev, Walmart's US technology manager, said that the extended version was able to post-order articles and plan events. It would also use computer vision to suggest recipes based on the content of a buyer's refrigerator. Walmart also develops an Associate Super Agent for branch employees and business employees. For sellers, suppliers and advertisers, a separate Marty agent is set up. The retailer also works on a developer superagent to test, create and launch future AI tools. Whether the agents would replace jobs, the company did not want to say. Dave Glick, Senior Vice President for Enterprise Business Systems, said that the tools would create new jobs without mentioning more details. (Photo by Growtika)

See also: OpenAI opens AI lab in Singapore while IMDA updates the AI framework

Would you like to know more about AI and Big Data from industry leaders? Check out the AI & Big Data Expo that takes place in Amsterdam, California and London. The comprehensive event is part of TechEx and takes place parallel to other leading technology events. Click here for more information. AI News is operated by TechForge Media. Discover more upcoming events and webinars on company technology.