AI governance becomes harder when an agent leaves the screen and enters a warehouse, delivery network or public space. A flawed software output can spread misinformation; a flawed embodied system can damage equipment, interrupt infrastructure or place a person in danger.
Singapore’s updated Model AI Governance Framework for Agentic AI offers a useful starting point. It treats agents as systems that plan, use tools and take multi-step actions, then links oversight to data access, autonomy, complexity and the reversibility of those actions.
The framework is not limited to robotics, but its emphasis on gradual deployment and continuous monitoring becomes especially important when software controls machines.
Physical consequences require deployment-based assurance
Embodied AI combines models with sensors, control software, semiconductors and mechanical systems. Each layer can fail independently, and unpredictable environments create conditions that were absent from laboratory testing.
Dr Ya-Qin Zhang of Tsinghua University argues that digital risks become amplified in the physical domain. Transport, drones, logistics, smart grids and critical infrastructure can all be affected as autonomy moves deeper into operations.
One-time certification cannot anticipate every interaction. Simulation and closed-course testing establish a baseline, but telemetry and post-deployment evaluation reveal the long tail of real-world failures.
Grab follows that sequence in pilots involving autonomous vehicles and delivery robots in Singapore’s Punggol district. It tests in simulation, then controlled and open environments, before expanding from a few machines toward larger fleets.
The key is not simply collecting logs. Operators need thresholds that identify unusual behaviour, procedures for intervention and an independent mechanism for taking a malfunctioning agent offline.
Human oversight must focus on significant moments
Continuous review of every autonomous action becomes impossible at scale. Singapore’s framework recommends human approval for high-stakes, irreversible or unusual actions while allowing bounded routine work to proceed.
Least-privilege access should apply to tools as well as data. A warehouse robot might read inventory and move approved goods within a zone without gaining permission to alter stock records or operate unrelated equipment.
Reversibility provides a practical guide. If an action can be cheaply undone and its outcome is easy to verify, more autonomy may be acceptable. Actions involving safety, financial transfer or critical permissions need stricter checkpoints.
Human supervision has its own failure modes. Automation bias can cause operators to accept a confident recommendation, while alert fatigue makes repeated warnings invisible. Override rates and response times help organisations evaluate whether oversight works in practice.
Users and employees should also understand what the system can access, which actions it can perform and what responsibility remains with them. Training must develop the professional judgement required to challenge automated behaviour.
Accountability stretches across a complex value chain
A physical AI deployment may involve a model provider, robotics manufacturer, chip supplier, integrator and site operator. Software updates and operational learning can further blur which party caused a later failure.
IMDA keeps organisations and people accountable even when the agent operates autonomously. Contracts and operating procedures therefore need named responsibilities for evaluation, updates, incident investigation and shutdown authority.
Semi-structured industrial environments may offer the earliest commercial path because conditions are more controllable. Chinese robotics company Galbot has deployed systems in retail, warehousing and pharmaceutical operations, while Japanese companies are most interested in manufacturing and dangerous tasks.
Purpose-built design will matter. Applied Materials argues that robotics depends on specialised sensors, efficient compute and packaging adapted to each industrial ecosystem rather than one universal machine.
Governance should follow the same principle. Controls must reflect the site, task and consequence rather than applying a generic policy to every robot.
Physical AI can address labour shortages and automate dangerous or repetitive work, but scale should follow evidence. Organisations need simulation, staged rollout, live telemetry, limited authority and a clear path to human control.
The decisive question is not whether an autonomous machine performs well during a demonstration. It is whether responsibility and safety remain intact after thousands of hours, multiple updates and the unexpected conditions that define the physical world.
