AI Agents Navigate Real-World Governance Challenges in Physical Environments

AI systems that can act on their own are starting to leave the digital world and enter real-world settings like warehouses, delivery routes, and public areas. This shift is raising questions about whether today’s AI regulations are equipped to handle systems that function in physical spaces.

So far, most AI governance efforts have centered on digital risks, such as biased outputs, false information, and harmful online content. But AI systems that interact with the physical world introduce new dangers, where malfunctions could damage property, disrupt infrastructure, or put people’s safety at risk.

On May 20, Singapore’s Infocomm Media Development Authority released version 1.5 of its Model AI Governance Framework for Agentic AI. The updated framework provides guidance for organizations rolling out AI agents capable of planning, making decisions, and carrying out multi-step actions to achieve goals set by users.

According to the framework, these agents can work with tools, external platforms, and other agents, including systems that modify databases, create files, manage devices, or execute transactions. It identifies access restrictions, oversight mechanisms, and human sign-off as key governance practices for deployment.

AI enters physical systems

At a recent AI summit in Singapore, conversations about robotics and embodied AI centered on operational safety concerns more typically linked to aviation, industrial equipment, and critical infrastructure management, rather than standard software regulation.

Participants also explored whether self-operating systems can function safely and consistently in unpredictable real-world conditions over long periods.

Dr. Ya-Qin Zhang, founding dean of the Institute for AI Industry Research at Tsinghua University, noted that embodied AI systems magnify risks already tied to autonomous software. He explained that failures could directly impact transportation networks, drones, supply chains, and essential infrastructure.

“Every risk found in the digital world grows larger in the physical world, and the physical world produces real-world consequences,” Zhang shared with MLex during the summit.

He further noted that vehicles, drones, smart grids, and other infrastructure could become vulnerable as AI systems become more deeply integrated into physical operations.

Discussions at the summit highlighted reliability, ongoing monitoring, and post-deployment verification as key governance priorities. The conversations pointed toward governance models based on simulation, real-time data tracking, and repeated testing, rather than relying solely on initial certification.

IMDA’s framework also advocates for phased rollouts, ongoing observation, and additional testing after deployment. It notes that agents interact dynamically with their surroundings, making it impossible to predict every risk before launch.

Monitoring shifts to a deployment concern

Grab, currently testing autonomous vehicles and delivery robots in Singapore’s Punggol district, emphasized that deployment governance relies heavily on simulation, testing, and continuous oversight.

“We run extensive simulations and conduct tests on both closed and open courses to ensure our robots perform reliably,” said Suthen Thomas Paradatheth, Grab’s chief technology officer, during a summit panel.

“Before expanding to hundreds of robots, we first validate everything in simulation and with a small number of units,” he added.

Grab also highlighted monitoring systems built to track robot behavior and catch unexpected failures after deployment.

“There’s a long tail of problems that can surface,” Paradatheth noted.

The IMDA framework advises organizations to evaluate agentic AI applications based on factors like data access, external system connectivity, level of autonomy, and task complexity. It also considers the scope and reversibility of agent actions, involvement of third parties, and overall system intricacy.

The framework further recommends restricting agent access to tools and systems, applying minimal necessary permissions, and establishing standard procedures for agent operations. Organizations should also have mechanisms to shut down agents when they malfunction.

Responsibility extends to more parties

MLex reported that embodied AI systems often involve multiple stakeholders across development, manufacturing, and deployment, including AI developers, robotics makers, chip suppliers, and infrastructure operators.

MLex also observed that assigning responsibility becomes more complex when systems continue evolving after deployment through software updates, telemetry, and operational feedback.

IMDA states that organizations and individuals remain responsible for agent actions, even when agents function independently. The framework calls for clearly defined accountability across the entire agentic AI value chain, from model and platform creators to deployers, tool providers, and end users.

Applied Materials noted that large-scale robotics deployment is closely connected to semiconductor economics and systems integration. Om Nalamasu, the company’s chief technology officer, explained that robotics systems will require improved sensors, better energy efficiency, advanced packaging, and specialized computing architectures.

Nalamasu added that robotics systems will need custom designs tailored to specific industrial settings rather than a universal solution for all environments.

Zhao Yuli, chief strategy officer of Chinese robotics startup Galbot, shared that Beijing is focusing on scaling deployment and industrial commercialization through government-supported testbeds, industry partnerships, and sustained funding programs.

Galbot has introduced humanoid robotics systems in retail, warehouse, and pharmaceutical operations across China, including stores that operate autonomously around the clock. Zhao indicated that semi-structured industrial settings are likely to serve as an early path to commercialization because they offer more predictable operating conditions.

Japan is concentrating more on establishing standards, building robotics datasets, and advancing safety governance. Professor Yutaka Matsuo from the University of Tokyo’s Graduate School of Engineering highlighted an “AI Association” initiative aimed at gathering 100,000 hours of robotics data to support robotic foundation models.

Matsuo also mentioned Japan’s AI Safety Institute and the Hiroshima AI Process as part of wider efforts to create governance standards for embodied AI systems alongside Singapore and other Asian nations.

Singapore outlines agent controls

Singapore’s framework identifies four governance areas for agentic AI: initial risk assessment, human accountability, technical safeguards, and end-user responsibility. The framework presents these as an ongoing cycle rather than a one-time evaluation.

The framework notes that human oversight must be adapted for agentic systems since continuously reviewing every workflow becomes unmanageable at scale. It recommends human approval at significant