How Dyna Robotics Hit 99.4% Success Over 24 Hours of Napkin Folding With Reward-Model Supervision

AI Engineer · youtube · 2026-09-24

Dyna Robotics co-founder and CTO Jason Ma explains how generalist robots go from "great video, bad product" to production reliability:

• 80–90% success makes a great video but a poor product; Dyna-1 folds restaurant napkins at 99.4% success over 24 hours straight, recovering even after pulling over the whole stack.
• The key is a reward model that watches and scores the robot's progress; score dips flag mistakes, letting the team collect targeted recovery data in a human-in-the-loop active-learning cycle.
• Data and architecture: a pre-training pyramid of 200,000+ hours; a reasoning model paired with a world action model.
• New tasks in under an hour of data; deployments in restaurants and a Sacramento laundromat, plus 3 days of shirt folding at CoRL with no site-specific data.

Original post →

More from Embodied

• HRI 2027 Adds Archival Industry White Paper Track for Real-World Robot Deployments — petitegeek · 2026-09-25
• Google dev imagines orchestrating AI agents via smart glasses: tap their shoulder, watch their screen — jason_mayes · 2026-09-25
• Climbing analysis in 3D with iPhone LiDAR: open-source SAM 3.1 + ViTPose demos — dosco · 2026-09-25
• Menlo open-sources Asimov 1 humanoid locomotion policy and Isaac Lab training code — freelerobot · 2026-09-25
• Human-Robot Dialogue workshop at IROS 2026: NVIDIA, MIT, Georgia Tech speakers lined up — dhadfieldmenell · 2026-09-25
• AI glasses shipments up 263% in H1; Zhiyuan delivers 20,000th robot — 创业邦 · 2026-09-25

Embodied channel →