The Final Frontier: Machine Learning Safety in Robotics With Paul Schmitt

Geoffrey Athey, Chief Executive Officer at Reynolds & Moore, joins Erik and Missy Reynolds, the founders of R&M, to announce that the company is becoming employee-owned. Geoffrey spent more than 20 years leading manufacturing operations before coming to R&M as Chief Operating Officer in 2024. In that role he turned the company’s tribal knowledge into processes built to scale. R&M has doubled in size 3 years in a row. Geoffrey explains how defining trust, transparency, and care as specific behaviors kept the culture intact through that growth. With ownership now added as a core value, Geoffrey makes the case for why employee-owned companies outperform their peers. An individual contributor owns 8 steps of a 100 step process, while an owner watches every handoff. For clients, Erik says, that means their problems become R&M’s problems.

Mark Nehrkorn, Senior Director of Functional Safety & Compliance at Symbotic, joins Erik Reynolds to talk about why certified parts don't guarantee a certified system. Over more than 30 years in machine safety, Mark has worked at Boston Dynamics and SICK. Today he leads the team keeping Symbotic's robotic warehouse systems safe. Every certified component arrives with its own conditions for use, and Erik points out that the more components a system combines, the more likely those conditions are to conflict. Mark explains when an off-the-shelf certified part makes sense, and when building thousands of robots makes it smarter to design your own. Safety work doesn't stop at commissioning. Symbotic's robots each drive more than 50 miles a day, while electronic components can reach end of life in years or even months. Mark describes how losing access to a single part can send a newly certified system back to the start. It's part of why he sees the traditional 20-year mission time as less relevant than it used to be. Mark Nehrkorn is also speaking at the R&M Physical AI & Safety Summit on October 13 at MassRobotics in Boston. Register for the Summit: https://reynolds-moore.com/events/rm-physical-ai-safety-summit/

Weihan Wang, Director at NexCOBOT, a NEXCOM International company, joins Erik Reynolds to talk about the hardware underneath robot safety. Weihan leads business development, product development, and functional safety for the robotics division and is the safety manager for the SCB 100. They cover how a motion controller, a PLC, and a safety controller differ, and what makes a safety rated PC a different product from a regular one. Weihan shares how NexCOBOT spent 4 years taking the SCB 100 through SIL 2 and PL d Category 3 certification. He also explains why force monitoring pushes robot makers away from safety PLCs and toward C and C++ on a PC platform, and how customers run redundant safety applications on 2 cores inside a single chip. The conversation then turns to humanoids, quadrupeds, and ISO 25785, which the ISO working group expects to publish around 2028. Weihan also describes the safety platform NexCOBOT is building on NVIDIA IGX Thor, where a safety island monitors the AI inference and a separate safety MCU serves as the backup. R&M and NexCOBOT are sharing a booth at NVIDIA GTC Berlin on October 20–22. Stop by to meet both teams.

Dr. Daniel Simon joins Erik on the podcast to talk about why the structure of safety critical software decides how safely it can be built and changed. Daniel is Distinguished Solutions Engineer at Qt Group, a member of the R&M Safety Partner Ecosystem. He co-originated Axivion Architecture Verification out of research at the University of Stuttgart and stayed with the work after Qt Group acquired Axivion. They cover what architecture verification checks, and why documented architectures go stale a few years into a project. Daniel explains why moving compute intensive code from CPU to GPU is an architectural exercise rather than a refactor. He also walks through the Axivion work on CUDA, including the collaboration with NVIDIA on their safety coding rules. The conversation ends on AI generated code and what it does to the person who has to sign off on it. Daniel is also speaking at the R&M Physical AI & Safety Summit on October 13 in Boston at MassRobotics, where he will cover why deterministic verification and emerging AI trust methods under IEC 61508 and ISO/IEC 22440 have to coexist as robotics teams pursue certification. The Summit is free to attend, and an evening reception at Qt Group follows. The next 2 days at MassRobotics, Reynolds & Moore is running Understanding ISO/IEC TS 22440, a training in collaboration with SafetyCert.AI. Register for the Summit: https://wkf.ms/4cdJRlF Register for the 2-Day Training: https://wkf.ms/4gBG3Mj

Chris Soranno, Manager of International Safety Standards at SICK Sensor Intelligence, joins Erik to take on a complaint the industry makes constantly, that safety standards are always behind the technology. Chris has spent 25 years inside that process. A technology has to prove it is not a passing trend before a committee will invest the effort, and by the time the first meeting happens, the market has already moved again. His argument is that a standard should say what has to be achieved, never how to achieve it. Chris points to the safety laser scanner, a proven technology that had nowhere to be applied because the application standards of the day named the light curtain as the only accepted answer. They also cover why risk assessment still holds up for machines whose behavior is harder to predict, why motion and contact are no longer treated as automatically hazardous, and what changes when the thing defeating your safeguard is an adversary rather than an accident.