Physical AI Safety

Safety in physical AI is categorically different from safety in digital AI. When a language model makes an error, the consequence is a bad response - harmful but reversible. When a physical AI system makes an error, a person can be injured, property can be destroyed, and the action cannot be undone. This irreversibility changes everything about how safety must be designed.

The Irreversibility Problem

In software AI, safety is primarily about filtering outputs before they reach users. In physical AI, the "output" is a physical action, and actions have consequences that cannot always be intercepted:

• A robot arm moving at full speed toward a human cannot be stopped if the safety system detects the hazard too late
• A self-driving car in a skid cannot "undo" that action
• A delivery drone that drops a package cannot retrieve it mid-fall

This means physical AI safety must be predictive (anticipate hazards before they occur) rather than purely reactive.

Defense in Depth

No single safety mechanism is sufficient. Physical AI systems use layered safety - multiple independent systems, each of which can halt or constrain the robot:

Layer 1: AI-level safety
  - LLM refuses unsafe instructions
  - Policy trained with safety constraints
  - Feasibility checking (SayCan-style grounding)

Layer 2: Perception-level safety
  - Human detection (stop if person enters zone)
  - Collision prediction (anticipate impacts)
  - Anomaly detection (unexpected situations)

Layer 3: Control-level safety (runs at 1kHz)
  - Joint torque limits (cannot exceed force limits)
  - Velocity limits (cannot move faster than safe speed)
  - Workspace limits (cannot move outside safe region)

Layer 4: Hardware safety
  - Emergency stop button (cuts power to actuators)
  - Compliant actuators (physically "give" on contact)
  - Fuses and circuit breakers

Each layer is independent - Layer 4 works even if
Layers 1-3 have software bugs.

ISO Safety Standards

Physical robot safety is governed by international standards, not optional best practices:

• ISO 10218 - Safety requirements for industrial robots. Defines requirements for fixed robot arms in manufacturing.
• ISO/TS 15066 - Collaborative robot (cobot) safety. Specifies force and speed limits for robots working alongside humans without physical barriers.
• ISO 13482 - Personal care robots. Covers mobile robots interacting with the general public (not just trained workers).
• ISO 26262 - Functional safety for automotive systems. Defines ASIL (Automotive Safety Integrity Levels) for AV components.

Power and Force Limiting

A core principle from ISO/TS 15066 for collaborative robots: limit the energy the robot can deliver to a human body to below injury thresholds. This is implemented via:

• Speed and separation monitoring - the robot slows as a human approaches
• Power and force limiting (PFL) - the robot's joint torques are limited so that any contact is non-injurious
• Hand guiding - the human can physically move the robot to teach new positions

Collaborative robots (cobots) like Universal Robots' UR series, Rethink Robotics' Baxter, and FANUC's CR series are certified to these standards and can operate without safety fences - a prerequisite for deployment alongside human workers.

Handling Uncertainty Safely

Physical AI systems must be safe not just when they are confident, but also when they are uncertain. Safe approaches to uncertainty:

• Conservative fallback - when the AI is uncertain, default to a safe action (stop, ask for clarification, return to home position)
• Uncertainty quantification - the system estimates its own confidence and refuses to act when confidence is below a threshold
• Human handoff - flag low-confidence situations for human review before proceeding
• Minimal footprint - prefer actions that are easy to reverse and have limited impact if wrong

Alignment in Physical Systems

AI alignment - ensuring AI systems pursue intended goals - has higher stakes in physical contexts:

• A reward misspecification that causes an LLM to produce unwanted text is annoying. The same misspecification in a robot reward function could cause the robot to pursue its objective in physically harmful ways (e.g., a robot rewarded for "efficiency" that ignores human safety to complete tasks faster).
• Physical AI systems need explicit "do not harm humans" constraints that override all other objectives - not just as training objectives, but as hard safety monitors at the control layer.
• The Asimov-inspired intuition (hardcode "do not harm humans") is the right direction, but the engineering implementation is non-trivial: the robot must recognize humans in all their variation, at all speeds, in all lighting conditions.