April 2025: a warehouse robot receives an ambiguous command. Its onboard Large Language Model interprets, makes a mistake, and strikes a worker. This is not science fiction, but an incident that the rush to commercialize physical AI risks multiplying. While companies push autonomous drones, self-driving vehicles, and industrial robotics to market, safety – functional safety engineering – remains stuck in paradigms designed for deterministic systems. It’s a gap that could consume billions and, worse, human lives.

A new risk: the LLM inside a robot’s body

Language models bring probabilistic logic into physical systems, a poor fit for traditional safety requirements. Deterministic software, when tested, produces the same output for the same input; an LLM, by contrast, introduces variation, hallucinations, and fragile contextual understanding. Imagine a robotic arm ordered to “move the box near the door”: the interpretation of “near” or the selection of the wrong box can trigger immediate physical harm. On edge devices, where VRAM is limited, quantization is often used to shrink the footprint, but this can further degrade the model’s reasoning. The safety gap arises precisely here: there are no standardized methods to certify a system whose response is not 100% reproducible, and typical language benchmark metrics (perplexity, task accuracy) do not measure physical risk.

Why edge computing isn’t enough (but is essential)

The very nature of physical AI mandates local or on-premise deployment: the latency of a cloud round-trip would render a drone unable to avoid an obstacle in flight. Yet on-premise infrastructure, however essential, does not automatically close the safety gap. Hardware constraints – power dissipation, available memory, total cost of ownership (TCO) – push toward smaller models, aggressive quantization, and reduced context windows. This trade-off between performance and reliability is a knot that those designing local deployments must untangle. The most promising approach is a hybrid architecture: an edge LLM for natural language interaction, paired with a deterministic safety controller that validates every physical action. However, such redundancy multiplies capital expenditure (CapEx) and integration complexity, driving away the very mid-sized companies that could benefit from intelligent robotics. AI-RADAR offers analytical frameworks on /llm-onpremise to weigh these trade-offs and understand when the additional safety investment becomes non-negotiable.

Regulation as a crutch, not a solution

The European AI Act classifies many AI systems interacting with the physical world as high-risk. Yet the technical standards for compliance are still being drafted, and competitive pressure may push products to market prematurely. The recent history of autonomous vehicles teaches us that regulation alone is not enough: a safety-by-design culture is needed, inspired by automotive processes (ISO 26262) but adapted to AI’s non-determinism. This demands transparency about training data, independent audits, and above all, the acceptance that an LLM in a physical system cannot be treated like ordinary software.

The knot to untie: trusting physical AI?

The safety gap is not merely technical but institutional and cultural. As LLMs become embedded in the real world, we will need unprecedented certification tools, specialized safety hardware (control co-processors, runtime monitoring), and a mindset that prioritizes human safety over speed. For IT decision-makers, the message is clear: on-premise and edge deployment guarantees low latency and data sovereignty, but it demands a safety architecture that goes far beyond the language model. Closing the gap means investing today in redundancy, testing, and skills, before headlines are written by avoidable accidents.