JB

Walden Robotics reduced robot failures by more than 10x by teaching its robots when to ask for human help. Their latest technical post describes an interesting approach to one of the biggest challenges in deploying robots in manufacturing: keeping production running while the robots are still learning. Walden trains its Large Behavior Models (LBMs) to recognize when they need assistance. Alongside the part of the model that generates robot actions, they’ve added a separate decision head that learns when to request help. When a robot encounters a situation it can’t handle, a remote operator can step in, resolve the problem, and hand control back. That intervention also becomes training data, helping the robot improve its task performance and learn when to ask for help next time. In one demonstration, Walden’s robots loaded camshafts into a mock machining conveyor for several hours without an uncaught failure, despite needing assistance multiple times. There’s an interesting connection to Toyota’s Jidoka principle, where machines detect abnormalities and alert human operators before problems affect production. What I’d be curious to see is how this performs across a larger fleet. How many robots can one remote operator reliably support? And how much does that ratio improve as the models get better? Those numbers would be particularly interesting for anyone thinking about the economics of deploying robots in factories. Worth reading: waldenrobotics.com/news/call… P.S. The selfie is from my visit to Walden in San Francisco a few weeks ago. Adrian was kind enough to invite us over and show us around. More footage from the visit coming in the next few weeks! —— Weekly robotics and AI insights. Subscribe free: 22astronauts.com