In Robotics Adoption, Safety Is Just as Important as Capability

When it comes to fundraising and newsmaking in robotics, safety typically takes a backseat to capability.

It’s rare for tech headlines to trumpet a robot clearing a new safety benchmark. And when investors are deciding where to put their money, they are more likely to be swayed by a dynamic new demo than by a detailed safety report.

But when the time comes to make robots commercially available, safety is arguably the more important factor. Legal teams, regulators and insurers won’t sign off on robotics systems until risks are clearly understood, quantified and mitigated. One serious safety incident is often enough to end a robotics initiative for good. Consider what has happened in the self-driving space, where Uber and General Motors both abandoned autonomous vehicle initiatives after high-profile accidents.

READ MORE: Are Humanoid Robots Ready for Industry? Hannover Messe Thinks So (Maybe)

In short, capability wins attention, but it’s safety that determines adoption. As they seek to bring systems to commercial markets, vendors must ensure that robots pass the following three safety levels:

Functional safety. Basically, robots need to be able to reliably follow hard constraints that prevent them from harming people or property. These are non-negotiable rules that robots must follow, no matter what other goals they are trying to accomplish. The idea is that physical safety should be both proactive and reactive, combining mechanisms that make certain unsafe actions impossible with the ability for robots to respond safely to dynamic environments.

Even this fundamental level of safety remains a challenge for robotics vendors, which is why commercially available robots in fields like manufacturing often still work in segregated work cells, rather than directly alongside humans. While there has been significant theoretical progress in enforcing hard constraints, many companies still struggle to reliably enforce these safety guarantees in real-world systems.

And as robotics systems become more advanced with the rise of new AI and machine learning concepts, they also become more complex.

Behavioral safety. At this level, safety extends beyond merely preventing physical harm, with vendors needing to ensure that robotic behavior is predictable and understandable. Some people refer to this as interpretability or explainability of AI. Really, it’s about trust and reliability: Can you reliably predict what the robot is going to do, or is there a risk of the system deviating from expected behavior in subtle or unexpected ways?

While the consequences of behavioral safety limitations may not be as dire as those of problems with functional safety, unpredictable behavior can still erode trust, disrupt operations and create risks for users. To illustrate behavioral safety, I sometimes compare robotics to air travel.

READ MORE: How to Speed Robotic Prototype Timelines without Sacrificing Quality

Flying 30,000 feet in the sky at 500 miles per hour should feel extremely dangerous. However, millions of people board airplanes every day and most without phobias take their seat without even thinking about what would happen if something went wrong. That’s because air travel is extremely predictable and reliable; accidents make the news because they are rare.

Similarly, behavioral safety allows humans to feel comfortable working alongside robots—not only because they present no physical danger, but also because they behave in ways that are consistent and predictable over time.

System-level safety. Beyond preventing robots from bumping into people and getting them to execute tasks reliably, we must seek to understand the overall systemic risks associated with introducing robots into businesses, worksites and homes. From an insurance perspective, for example, this means quantifying both the probability of failures and the severity of impacts, then translating those risks into real-world costs.

While this sounds simple, it’s actually an extraordinary challenge because we have so little data about how robots behave in real-world environments alongside humans over extended periods. The use of robots around people has been relatively limited (largely because of limitations around functional and behavioral safety), and so there has been little opportunity to develop robust methodologies to measure system-level safety.

READ MORE: Adaptive Motion Systems: Building Machines That Learn

In the absence of better tools, companies may rely on brute-force testing, running systems repeatedly in the real world in hopes of surfacing rare failures. However, this approach is both costly and inefficient, and it still may not uncover low-probability—but important—problems.

Safety isn’t only less exciting than capability. In some ways, it is also a much more challenging problem to solve. Getting a robot to perform a task in a demo, after all, is far easier than ensuring the robot can perform that task safely, predictably and reliably across all scenarios, while also quantifying the risks and consequences of failure.

However, this work is essential. Without it, robots will largely remain confined to controlled environments like labs and work cells—rather than making their way into the complex, human-centered settings where they can have the greatest real-world impact.