RI Seminar: Ben Burchfiel : Learning in Production: Building Physical AI at Walden Robotics
Ben Burchfiel, co-founder and CTO of Walden Robotics, discusses building physical AI that learns from factory work at a Carnegie Mellon Robotics Institute seminar.
In this talk
- 1:16
Manufacturing as a learning flywheel
Walden aims to build robots that take on physical work, starting in manufacturing. Burchfiel argues that robots helping build and maintain other robots can expand the fleet, increase the variety of work it encounters, and generate experience that improves its AI. The first platform is a wheeled mobile manipulator, which he calls the zeroth-generation robot.
- 9:13
Deployment feedback and human assistance
Burchfiel says deploying a model is only the start; improvement depends on bringing field experience back into the AI, hardware, and product. At customer sites, Walden uses remote assistance so a robot can pause when it needs help, receive a brief correction, and resume autonomy. He describes interventions lasting roughly two to eight seconds, which also provide targeted training data while supporting production work.
- 17:20
Larger models learn faster on the job
Walden’s models combine simulation, first-person video, UMI, data from other robot bodies, and fleet experience, with deployment data grounding those sources. For its latest pretraining, the company used about 100,000 hours of general data plus thousands of hours of internal deployment data. Burchfiel says the goal was five times faster learning at a customer site, and that the team met its target of reaching a week’s prior performance by the end of the first shift.
- 20:10
Field results and unexpected capabilities
Walden tracks autonomy and mean time between interventions across customer processes, using deployment as an ongoing evaluation signal. Burchfiel reports that successive model generations improved and pushed back apparent performance plateaus; stronger models could also slow down under distribution shift, prompting the team to track task speed as well. In unscripted demonstrations, the model sometimes kept distance from people, moved obstacles, or recovered after a dropped grasp, though he says these behaviors were uneven and not explicitly trained as specific tasks.
- 45:33
Simulation, hardware, and scaling the fleet
Burchfiel says simulation has become more predictive of production performance as models grow stronger, and Walden runs diverse simulation evaluations during training to catch regressions. The company also reconstructs real environments in simulation, including an office kitchen, to probe behaviors in new settings. He sees fleet growth and deployment data as major opportunities and bottlenecks, while noting that durable hardware and efficient edge inference remain practical constraints.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.