Robots with AI Raise Safety and Control Challenges

Regulators and companies are confronting new risks as autonomous artificial intelligence is integrated into robots, sensors and industrial equipment. Google DeepMind has introduced Gemini Robotics, Gemini Robotics‑ER and a safety evaluation dataset called ASIMOV to test embodied AI behavior and reasoning.

Industrial automation provides concrete data for the debate. The International Federation of Robotics reported 542,000 industrial robots were installed worldwide in 2024. The federation projects 575,000 installations in 2025 and more than 700,000 by 2028. Market researchers applying the label “Physical AI” to robotics, edge computing and autonomous machines estimate the global market at $81.64 billion in 2025 with a projection to $960.38 billion by 2033.

Governance for physical systems differs from software-only automation because model outputs can become robot motion, machine instructions or operational decisions based on sensor readings. Safety must be engineered into hardware and software to limit force, prevent collisions and maintain stability. Systems need higher-level checks to assess whether a requested action is appropriate, to detect task success, to decide whether to retry or stop, and to record activity for review.

DeepMind first introduced Gemini Robotics and Gemini Robotics‑ER in March 2025. Gemini Robotics is a vision-language-action model that accepts natural-language commands for manipulation tasks. Gemini Robotics‑ER focuses on embodied reasoning, spatial logic and multi-step task planning. An April 2026 update brought Gemini Robotics‑ER 1.6, which adds stronger spatial reasoning, planning through intermediate steps and improved success detection. These models are available in preview through the Gemini API and developers can use Google AI Studio to build agentic applications.

DeepMind lists partnerships and trusted testers that include Apptronik, Agile Robots, Agility Robotics, Boston Dynamics and Enchanted Tools. Collaborative work has covered tasks such as instrument reading that require visual interpretation and reliable assessment of physical conditions.

Controls familiar in enterprise AI-access rights, audit trails, refusal behavior and escalation paths-become harder to manage when models connect to hardware. A 2026 industry study found about one-third of organizations reported relatively mature approaches to strategy, governance and agentic AI oversight. Standards such as the NIST AI Risk Management Framework and ISO/IEC 42001 offer governance structures that account for model behavior, connected machines and the operating environment.

Applications span industrial inspection, manufacturing, logistics, facilities and warehouses. Regulators and firms must decide what data systems may access, which tools and actions require human approval, how activity is logged and how safe operation is verified before granting autonomous authority.