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

Why watch

Jason Ma explains how Dyna Robotics achieved 99.4% success rate with their Dyna-1 model folding restaurant napkins for 24 hours straight by using a reward model that detects mistakes and triggers targeted data collection for fine-tuning. The talk covers their pre-training data pyramid of 200,000+ hours, their reasoning model paired with a world action model architecture, and why 80-90% demo success rates fail as products. Target audience: robotics engineers, ML researchers, and product managers working on real-world robotic deployments.