How Edge Computing Powers Real-Time Cobot Safety Systems

Lucas: There’s this number that sticks in my head from a safety engineering paper last year: a collaborative robot arm moving at two meters per second needs to detect a human and adjust its trajectory in under 10 milliseconds to avoid a serious impact. Luna: Ten milliseconds — that’s roughly the latency of a single frame at 100 frames per second. So cloud is basically off the table. Lucas: Exactly. By the time a signal goes from the cobot’s torque sensor to a cloud server and back — 50, 80, even 100 milliseconds if the network’s congested — the arm has already moved twenty centimeters. That’s the difference between a bump and a bruise, or worse. Luna: So edge computing isn’t a nice to have here. It’s a safety requirement. Lucas: That’s the angle. And I want to anchor on a specific deployment that I think illustrates the trade-offs really well. BMW’s assembly plant in Regensburg, Germany, has been running collaborative robots — cobots — alongside human workers for about three years now. But in 2025, they upgraded the safety system to an edge-based architecture. Luna: What did they have before? Lucas: A mix of light curtains and pressure-sensitive floor mats. Those are passive safety measures — if a human crosses a boundary, the robot stops. But stopping and restarting a production line costs about 20 seconds of downtime per event. In a plant making roughly 1,000 cars per day, those 20 seconds add up fast. The edge approach lets the robot keep moving but adjust its torque and speed dynamically. Luna: So instead of a hard stop, the cobot slows down or changes path in real time. That’s a huge productivity gain. Lucas: Right. The Regensburg setup uses 48 cobots across two assembly cells, all sharing a single edge node — a custom server with an NVIDIA Jetson module and a dedicated FPGA for sensor fusion. The cobots stream torque, position, and proximity data at 1 kilohertz. That’s 1,000 data points per second per robot. Luna: Forty-eight robots at 1 kilohertz each — that’s 48,000 data points per second hitting that one edge node. How does it not saturate? Lucas: That’s where the FPGA comes in. It pre-filters the data — discards redundant readings, only passes events that cross a threshold. The Jetson runs a lightweight vision model that detects human joint positions from overhead cameras. The model is only 500 kilobytes. It runs inference in 3 milliseconds. Luna: Three milliseconds for a full human pose estimate. That’s impressive. What’s the accuracy? Lucas: BMW published results showing 94.2 percent accuracy on their test set — which is good enough for torque modulation but not for a full stop decision. They still have a hard stop fallback if the model confidence drops below 80 percent. Luna: So the edge is doing a probabilistic safety check, not a binary one. That’s a different mindset from traditional industrial safety where everything is hardwired relays. Lucas: Exactly. And that shift opens up a really interesting debate. Some safety engineers argue that probabilistic models have no place in life-critical systems. But the counterargument is that the hardwired approach forces a stop, which introduces its own risk — workers rushing to restart, bypassing guards. Edge-based adaptive safety keeps the human in the loop without the productivity penalty. Luna: What about cost? A Jetson plus FPGA module isn’t cheap. How does that compare to the old light curtain setup? Lucas: BMW’s internal numbers say the edge node cost about $8,000 per cell — that’s for the hardware, integration, and the custom model training. The old light curtain system cost around $2,500 per cell. But the edge system paid for itself in 14 months because of reduced downtime. Each cell saves about 18 minutes of lost production per shift. Luna: Eighteen minutes per shift, two shifts per day, 250 working days a year — that’s 150 hours of saved downtime annually. At a plant that values a minute of production at roughly $100, that’s $900,000 per year per cell. Lucas: You did the math. And that’s just one plant. BMW has 30 plants worldwide. They’re now rolling this out to four more in 2026. Luna: So the business case is solid. But I want to push back on the edge vs. cloud framing. Couldn’t you run that vision model on a 5G-connected cloud with a dedicated slice and get under 10 milliseconds? Lucas: Good question. A 5G ultra-reliable low-latency slice can get you down to 5 milliseconds one-way — so round trip is 10 milliseconds. But that’s best-case, with no contention. In a factory with hundreds of devices, the actual measured latency at BMW’s Regensburg plant was 14 to 22 milliseconds during peak production. Plus you have the jitter problem. The edge node delivers consistent 3-millisecond inference every single time. Luna: Jitter kills real-time control. If your latency varies by 10 milliseconds, you can’t guarantee safety. Lucas: That’s the key insight. Determinism matters more than raw speed in safety systems. And edge gives you that determinism because you control the entire pipeline. Luna: One more thing — what about model updates? If the edge model is only 94 percent accurate, do they retrain it periodically? Lucas: They do. BMW retrains the model every two weeks using data collected from the edge nodes. The retraining happens on a central server, and the new model weights — about 2 megabytes — are pushed to the edge nodes during a scheduled downtime window. The whole update takes less than 30 seconds. Luna: So the edge is handling inference, but training still happens centrally. That’s a common pattern. Lucas: Right. And it’s worth noting that some automakers are now building their own custom edge hardware. Toyota, for instance, has a project called ‘Edge Brain’ that uses a custom ASIC instead of an off-the-shelf Jetson. They claim 1.2 millisecond inference time with 97 percent accuracy. Luna: Custom ASICs are expensive to develop, but at Toyota’s scale it probably makes sense. Especially if they can reuse the chip across multiple plants and even in vehicles. Lucas: That’s exactly their plan. They’re designing the chip to work both for cobot safety in factories and for pedestrian detection in cars. Same core architecture, different model on top. Luna: That’s a smart long-term play. And it shows how edge computing for cobot safety is really a subset of a larger trend — embedding AI into physical systems where latency is non-negotiable. Lucas: Yeah. And you know, conversations like this — where we dig into the real engineering trade-offs — that’s exactly the kind of content that keeps me excited about this show. If you find value in it too, and you want to help us keep it ad-free and independent, there’s a simple way to support the work. You can buy us a coffee at buy me a coffee dot com slash fexingo. It’s just a small gesture, but it genuinely helps us keep the lights on. Luna: And we don’t run ads, so listener support is what makes episodes like this possible. Even a few dollars makes a difference. Lucas: So back to the edge — I think the bigger takeaway is that cobot safety is forcing a rethinking of how we design industrial control systems. The old model was: stop everything and ask questions later. The new model is: sense continuously, decide locally, adapt instantly. Luna: And that shift is only possible because edge computing finally makes the latency budget feasible. It’s a quiet revolution, but it’s happening on factory floors right now. Lucas: Quiet revolution — I like that. Next time you’re in a plant or even a warehouse, look at the robots. They might be watching you more closely than you think.