Runtime Safety Filtering for Two-Terminal Hazards in Robotic Battery Recycling

Robots are learning to do delicate work, like sorting and recycling the batteries that power our modern world. To teach them, researchers often use artificial intelligence that learns by watching videos of humans performing tasks. These systems, known as vision-language-action models, can figure out how to pick up a battery or move a tool just by looking at a scene and understanding a simple instruction. However, there is a gap between what these smart systems know and what they can safely do. A robot might understand that it needs to move a battery from point A to point B, but it might not realize that moving it too close to a specific metal part could cause a dangerous electrical short. To bridge this gap, engineers use a "safety filter." Think of this filter as a vigilant supervisor standing between the robot's brain and its muscles. The supervisor watches every move the robot plans to make. If a move looks like it will break a safety rule, the supervisor stops it and tells the robot to do something else instead. The challenge is figuring out exactly what counts as a dangerous move and deciding what the robot should do when it is told to stop.

In the specific case of recycling batteries, the danger is unique. A battery cell has two metal terminals, one positive and one negative. If a piece of metal, like the battery casing itself or a tool, touches both terminals at the same time, it creates a bridge that allows electricity to flow where it shouldn't, causing a short circuit. This is different from a standard collision, where hitting a single object is the problem. Here, the danger only exists if the robot gets close to both terminals simultaneously. A robot could be very close to one terminal and be perfectly safe, but the moment it approaches the second one while still near the first, the risk appears. Researchers at the National University of Singapore and several Australian universities set out to test how best to program these safety supervisors to handle this specific "two-terminal" hazard. They wanted to know if the way the danger is defined matters more than what the robot does when it is stopped.

The team built a simulated environment using a standard robotics testbed called LIBERO, where a robot arm had to move a grocery item, which acted as a stand-in for a conductive battery part, across a workbench. They programmed the robot with a frozen AI policy, meaning the robot's decision-making brain was fixed and could not learn new tricks during the test. The researchers then tested different versions of the safety filter. First, they tried different ways of defining the "danger zone." One version used a simple rule: stay away from either terminal. Another used a more complex rule: stay away from the area where you could touch both terminals at once. They also tested a hybrid approach that combined both rules. Crucially, they did not just test these rules at one fixed distance. They adjusted the "margin," or the buffer zone, for each rule to see how much space was needed to keep the robot safe without stopping it too often.

What they found was surprising. When they adjusted the buffer zones for each rule to find the best possible balance between safety and getting the job done, all three definitions of the danger zone performed almost exactly the same. Whether the supervisor was looking for a simple "stay away from one" rule or a complex "stay away from both" rule, the results were nearly identical once the distance was tuned correctly. This suggests that the specific shape of the safety rule is less important than getting the distance right. The real difference came from what the robot did when the supervisor said "stop."

The researchers tested four different ways for the robot to react when a move was rejected. The first was to simply hold still, freezing the robot in place. The second was to "retreat," moving the robot backward away from the danger. The third was to "sample search," where the robot tried a few slightly different moves to see if one was safe. The fourth was a complex mathematical projection that tried to find the safest possible move that was still very close to what the robot originally wanted to do. The results showed that the "hold still" strategy was the worst. It stopped the robot from finishing its task about 30 percent more often than the "retreat" strategy, without actually making the environment any safer. The robot just sat there, failing to complete the job. On the other hand, the strategies that tried to be minimally invasive, like the mathematical projection or the sample search, allowed the robot to finish its task more often, but they left a much higher risk of the robot getting too close to the danger zone. The "retreat" strategy, which moved the robot clearly away from the hazard, offered the best balance: it kept the robot safe and allowed it to finish the task almost as often as the unfiltered robot would have.

The study also looked at how these filters behaved when the robot made mistakes or when the environment changed. They tested the system with a different robot policy and different battery shapes, and the "retreat" strategy remained the most effective. They even simulated errors in how the robot measured distances. They found that if the robot was wrong about the size of the object it was holding, the safety filter became much less effective than if it was just slightly wrong about where the terminals were. This highlights that knowing the size of the payload is just as critical as knowing where the hazards are. Furthermore, they discovered that a safety system that only checks for physical contact would fail completely. Many of the dangerous moments happened when the object was still a few millimeters away from the terminals, a distance too small to touch but large enough to cause a short. A contact-based monitor would miss these events entirely.

Ultimately, this research shows that for robots handling hazardous materials, the most important part of the safety system is not necessarily the complex math used to define the danger, but rather the simple, clear decision of what to do when the danger is detected. A supervisor that tells a robot to back away is far more effective than one that tells it to freeze or one that tries to nudge it just slightly out of the way. The study, conducted entirely in a high-fidelity simulation, provides a clear roadmap for building safer robots. It suggests that engineers should focus less on inventing new, complex geometric rules for danger and more on choosing a robust, clear reaction strategy that moves the robot away from risk while still allowing it to complete its work. As the world moves toward automating the recycling of millions of batteries, these findings offer a practical guide for ensuring that the machines doing the work do not accidentally cause the very hazards they are meant to manage.