Embodied Intelligence: Safety Beyond the Screen
When AI leaves the browser and enters physical and virtual space — robotics, autonomous systems, extended reality, virtual humans — a new class of safety questions follows it.
Summary
The bottom line for decision-makers
Embodied AI does not merely add new applications to the existing safety agenda; it changes the shape of the agenda itself. A system that acts in the world operates in a closed loop with an environment it can only partially sense, under time budgets measured in milliseconds, with outcomes that frequently cannot be reversed. Safety in this regime cannot be certified once and assumed thereafter. It must be engineered into the control loop, bounded by mechanisms that remain safe even when the learned components fail, validated across the gap between simulation and reality, and governed by accountability structures that can answer for physical harm. This report supplies a survey of the risks, a taxonomy of the failure modes, and a set of design principles to guide the field.
When AI leaves the browser and enters physical and virtual space — robotics, autonomous systems, extended reality, virtual humans — a new class of safety questions follows it.
Key findings
- Bound the learned system with an unlearned one. Wrap probabilistic, learned controllers inside a verified safety envelope — a supervisory layer that can guarantee constraints such as collision avoidance and force limits regardless of what the learned policy proposes. Never let a neural network be the last line of defense.
- Treat the sim-to-real gap as a first-class hazard. Simulation is indispensable for embodied testing, but the discrepancy between simulated and physical behavior is itself a source of failure. Quantify it, stress it, and stage deployment so that reality is met gradually and under supervision.
How to cite this report
Embodied Intelligence: Safety Beyond the Screen. FAIR Labs Research Report. Fair Artificial Intelligence Research Labs, 2025. Available at https://fairlabs.ai/research/embodied-intelligence-safety
FAIR Labs (Fair Artificial Intelligence Research Labs) is a nonpartisan 501(c)(3) research institute in Washington, DC. Its reports may be reproduced with attribution for non-commercial purposes. For interviews, briefings or data requests, use the contact form.