Dyna chief scientist says its models folded restaurant napkins at 99.4% success over 24 hours

A success rate is a ratio, and the page reports one side of it. At 99.4% the failure rate is 0.6% ^[10], or six failed folds in every 1,000 napkins ^[11]. Latent Space's summary of the talk does not say how many napkins the 24-hour run covered, what counted as a failure, or whether a person ever reset the table.

The one operational detail is the recovery. According to the summary, the run included the robot recovering after it pulled over the whole stack ^[13]. If that recovery counts inside the 24 hours, the test is harder than a clean repeat loop. The episode's topic list ties agentic harnesses to multi-step, hour-long physical workflows ^[7]. Which layer produced this particular recovery is open.

The 99.4% figure and the Din Tai Fung deployment sit in different sentences on the page. The figure belongs to a 24-hour run ^[12]. The restaurant is where Dyna's robots fold thousands of napkins daily ^[14]. If the same rate held there, every 1,000 napkins would still include about six failures ^[11].

Latent Space wrote that Dyna-2.1 "has to be seen to be believed" ^[6]. The evidence on offer is that demo and the account of co-founder and Chief Scientist Jason Ma ^[8].

Ma frames the real challenge in robotics as reliable, long-horizon autonomy, per the page's summary of his AIE talk ^[3]. A 24-hour run on one task is a test of that. It cannot be read as a measure across tasks, and one company's own figure from one task can only speak for Dyna-2.1. In our view the Din Tai Fung deployment is the better test, because it accumulates failures over weeks instead of hours.

Dyna names its own open problems on the same page: dexterous hands still overheat, and missing sensory data remain major bottlenecks ^[5].