Variation Kills Automation: Teaching Robots Stiffness | Sze Cheong & Elijah Ong, Devol Robots

Walk into a factory that already owns robots and you'll find the end of the line automated: packaging, palletizing. Everything in between is still done by hand - plugging connectors, assembling parts, moving a lens into a jig with 10 to 20 microns of tolerance. As Sze puts it, variation kills automation. Sze Yuan Cheong (Co-Founder & CEO, Devol Robots - a decade running manufacturing businesses before he touched a robot) and Elijah Ong (Co-Founder & CTO - turned down a Stanford PhD to help build a force-controlled arm from scratch at a three-person startup in Austin) argue that robot AI is scaling in the wrong direction. A camera can't tell a robot arm resting on a table from one pressing into it with 50 newtons. Devol teaches robots stiffness and damping, treats robot data as living in curved space instead of flat vectors, and pretrains on about 1,000 hours of data. In their own benchmark the model succeeds about 90% of the time on contact-rich tasks, where two open-source baselines land between 30 and 40%. We cover what jigs, fixtures, and integrator hours cost a factory, the keys-in-your-pocket test for what vision can't capture, an Ethernet plug insertion phase by phase, what happens when a manufacturer calls with a new task (one demonstration, then about an hour of the robot teaching itself), whether the VLA labs can scale their way to the same place, how the model runs on position-controlled robots, where it breaks, and the hard lesson from trying to build an entire hardware stack in house. Chapters: 00:00 Why the company is named after George Devol 00:50 Variation kills automation: what factories still do by hand 02:28 The hidden cost of jigs, fixtures, and integrator hours 04:54 Inside an optics line: 20 jigs and 10 microns of tolerance 07:22 Why a camera can't see 50 newtons: stiffness and damping explained 09:18 Finding your keys in your pocket without looking 10:03 Turning down a Stanford PhD to build a force-controlled arm 11:49 Series elastic actuators and a 3 a.m. phone call 12:46 Robot data lives in curved space 15:28 The Ethernet plug, phase by phase 17:09 Why 1,000 hours of pretraining is the floor 18:59 20 demos vs. 100: the benchmark against VLA baselines 19:52 A new task on the factory floor: one demo, then an hour of robot self-play 21:59 The bitter lesson: can the VLA labs scale their way here? 23:34 The classical robotics lesson AI forgot: find the right abstraction 24:40 Running on position-controlled robots 26:32 Where the model breaks: threading a needle 27:38 Hot take: stop scaling from pixels 28:41 Devol One: blending VLA, JEPA, and world action models 31:20 Five years out: AI rediscovers classical robotics 32:11 Slow brain, fast brain, and why low-level control matters 33:18 Nobody is working on the execution problem 34:05 Advice for founders: don't build everything yourself 35:40 Where to find Devol Learn more about Devol Robots: https://www.devolrobots.ai The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Hosted by Bogdan Cristei of OPTIM VC. Subscribe to the newsletter: https://www.optim.vc