Robot Demos Are Easy. Reliability Is Hard — Jason Ma, Dyna Robotics

Jason Ma of Dyna Robotics presents their approach to achieving high reliability in robotic manipulation, moving beyond impressive demos to product-ready performance. The core innovation is a reward model that monitors robot progress, enabling the collection of targeted recovery data for human-in-the-loop active learning and fine-tuning. This process is crucial for overcoming the limitations of standard post-training methods that stall at lower success rates. The talk also covers Dyna's extensive pre-training data pyramid (200,000+ hours) and their architecture, which pairs a reasoning model with a world action model. Case studies include 24-hour napkin folding with 99.4% success, T-shirt folding for three days, and opening Red Bull cans at live events, all achieved with minimal or no site-specific data.