The Ethical Dilemma of Physical AI in Autonomous Systems

Governance Challenges in Physical AI: A Deep Dive

The landscape of governance surrounding Physical AI is evolving rapidly as autonomous AI systems find their way into robots, sensors, and industrial equipment. Beyond the simple completion of tasks, the crux of the issue lies in how the actions of these AI agents are tested, monitored, and halted when they interact with real-world systems.

The realm of industrial robotics serves as a robust foundation for this ongoing discourse. According to the International Federation of Robotics, the global installation of industrial robots reached a staggering 542,000 units in 2024, more than doubling the figures from a decade prior. Projections indicate further growth, with installations expected to surpass 575,000 units in 2025 and soar past 700,000 units by 2028.

Market analysts are broadening the scope of Physical AI to encompass a wider array of systems, including robotics, edge computing, and autonomous machines. Grand View Research estimates the global Physical AI market to be valued at US$81.64 billion by 2025, with projections soaring to US$960.38 billion by 2033, contingent on the diverse interpretations of intelligence embedded in physical systems.

The governance challenge posed by Physical AI differs significantly from that of software-only automation, given the physical presence of these systems in workplaces, infrastructure, and amid human users. The ability of these systems to be linked to equipment necessitates clear safety boundaries. A mere model output can translate into a robot’s movement or a machine’s instruction, or even evolve into a decision grounded in sensor data, underscoring the critical importance of safety limits and escalation pathways within system design.

A notable case in point is Google DeepMind’s foray into the realm of robotics, exemplified by the introduction of Gemini Robotics and Gemini Robotics-ER in March 2025. These models, built on Gemini 2.0 for robotics and embodied AI, showcase a vision-language-action model (Gemini Robotics) tailored to directly control robots, and a counterpart (Gemini Robotics-ER) focusing on embodied reasoning encompassing spatial comprehension and task strategizing.

The utilization of such models by robots necessitates the identification of objects, comprehension of instructions, sequencing of movements, and evaluation of task completion accuracy. This poses a dual challenge entailing model behavior and the mechanical constraints of the system, accentuating the need for a meticulous control strategy.

Google DeepMind emphasizes the indispensable traits of generality, interactivity, and dexterity in crafting efficient robots. Generality encompasses the handling of unfamiliar objects and environments, while interactivity pertains to human interaction and dynamic conditions. Dexterity, on the other hand, underscores the mastery of physically demanding tasks requiring precision.

Gemini Robotics, as highlighted in Google DeepMind’s launch materials, exhibits the capacity to interpret natural-language instructions and execute multi-step manipulation tasks such as folding paper, packing items into a bag, and manipulating objects unseen during training, thereby showcasing the multifaceted technical requirements integral to Physical AI.

Beyond language comprehension, systems under the ambit of Physical AI necessitate visual perception, spatial reasoning, task planning, and success detection capabilities. In the domain of robotics, the ability to discern task success is pivotal, as the system must determine whether a task has been accomplished, whether a retry is warranted, or if cessation is imperative.

The introduction of Gemini Robotics-ER 1.6 by Google DeepMind in April 2026 underscores the evolution of these functions in advanced models, boasting spatial logic, task planning, success detection, and the ability to navigate through intermediate steps while deciding on progression or retrial.

Google’s developer ecosystem, comprising Google AI Studio and the Gemini API, facilitates the integration of Gemini models into applications, thereby streamlining the testing and deployment of agentic applications in the realm of embodied AI.

Navigating the governance maze in Physical AI becomes increasingly intricate when these systems exhibit the capability to summon tools, generate code, or trigger actions. Robust control measures are imperative to delineate data access, tool usage, human approval prerequisites for actions, and activity logging for retrospective review.

McKinsey’s 2026 AI trust research underscores the imperative to fortify governance frameworks in enterprise AI, especially as AI systems assume greater autonomy. In the realm of robotics, safety considerations extend beyond the behavioral facets of the machine to encompass lower-level controls like collision avoidance, force thresholds, stability parameters, and higher-order reasoning pertaining to the safety of requested actions within the prevailing context.

Introducing ASIMOV, a dataset aimed at evaluating semantic safety in robotics and embodied AI, Google DeepMind endeavors to assess the systems’ comprehension of safety-oriented instructions and their ability to avert unsafe behaviors in physical settings, underscoring the criticality of robust safety controls in Physical AI.

The intricate orchestration of controls, from access rights to audit trails, refusal behavior, escalation pathways, and testing, poses a formidable challenge when dealing with systems interconnected with robots, sensors, or industrial equipment. Governance frameworks like the NIST AI Risk Management Framework and ISO/IEC 42001 offer structured pathways to mitigate AI risks and delineate responsibilities across the system lifecycle, necessitating a tailored approach in the context of Physical AI.

Google DeepMind’s collaborative efforts with robotics companies exemplify its commitment to advancing embodied AI, with strategic partnerships fostering innovation in humanoid robots and robotics tasks requiring visual acumen, task planning acumen, and reliable assessment of physical conditions, thereby contributing to the maturation of Physical AI applications.

The applications of Physical AI extend across industrial domains encompassing inspection, manufacturing, logistics, facilities, and warehouses, necessitating systems to interpret real-world conditions and operate within predefined limits. The core governance query revolves around the establishment of these limits prior to enabling autonomous systems to make or execute decisions.

In conclusion, the dynamic landscape of Physical AI presents a plethora of governance challenges that necessitate a nuanced approach to balance innovation with safety, efficacy, and ethical considerations. The collaborative efforts between industry stakeholders and AI pioneers like Google DeepMind are instrumental in shaping the contours of Physical AI governance, paving the way for a future where autonomous systems coexist harmoniously with human operators in diverse real-world settings.

For more insights on AI and big data trends from industry thought leaders, be sure to explore the upcoming AI & Big Data Expo events in Amsterdam, California, and London, featuring cutting-edge discussions on the latest technological advancements and industry best practices. Stay ahead of the curve with TechForge Media, your go-to source for enterprise technology events and webinars.

(Photo by Mitchell Luo)

Discover more: AI agent governance takes focus as regulators flag control gaps

This article is powered by TechForge Media. Explore a plethora of upcoming enterprise technology events and webinars here.