Robot Demos Are Easy. Reliability Is Hard

Jason Ma, Dyna Robotics26:42 · Sept 2026 · 10K views
Thumbnail for Robot Demos Are Easy. Reliability Is Hard Watch on YouTube
TL;DR
  1. 1

    Dyna Robotics combines broad robot pre-training with targeted active learning to make generalist manipulation policies reliable enough for commercial work.

  2. 2

    Its reward model watches task progress, detects dips that signal mistakes, and helps the team collect focused recovery data instead of manually reviewing every run.

  3. 3

    Dyna-1 folded restaurant napkins with a reported 99.4% success rate over 24 hours and later folded T-shirts for three days at CoRL 2025 without site-specific data.

Summary

Jason Ma explains how Dyna Robotics is trying to move general-purpose robots beyond impressive demos. Models that succeed 80 to 90% of the time are still poor products because repeated tasks quickly accumulate failures. Dyna starts with a broad pre-training data pyramid containing more than 200,000 hours from human video, public datasets, simulation, robot tasks, and deployment sites. Its architecture combines a reasoning model with a low-level world action model for precise physical control. For hard tasks such as restaurant napkin folding, standard post-training reached about 80% success and then stalled when mistakes caused the robot to leave its training distribution. Dyna added a reward model that scores progress through a task. Researchers use reward dips to find failures, collect targeted recovery examples, and fine-tune again. Ma also describes deployments in restaurants and a laundromat, plus a T-shirt-folding demonstration at CoRL 2025 with no new site data.

Key ideas
01:12

Dyna uses deployments to decide which robotics problems deserve research effort

Ma describes a research and deployment flywheel. Dyna develops research and hardware internally, then deploys robots at customer sites to learn which tasks and workflows have commercial value. The deployments provide high-quality data from environments outside the lab and reveal what the models and hardware still cannot do. He says Dyna had more than five deployment sites at the time of the talk. This feedback helps the team avoid choosing problems only because they make attractive videos. It also connects product work to research priorities, since the failures seen in restaurants, laundromats, and other sites point to specific gaps in the robot's capabilities.

02:16

The robot foundation model combines broad data with separate levels of control

Dyna's model takes camera observations of the world and produces robot joint positions or torque commands. Ma says the training data follows a pyramid. The broad base includes human video, public datasets, and some simulation data. Robot-collected data adds diverse industrial, household, laundromat, and hotel tasks. Deployment data helps close the gap between laboratory training and customer environments. The training pipeline contains more than 200,000 hours of data. The architecture pairs a high-level reasoning model with a low-level world action model that produces fine-grained, high-frequency actions. The reasoning side handles semantic understanding, while the action side handles physical interaction and recovery.

05:01

A strong pre-trained model can learn unfamiliar tasks from less than an hour of examples

Ma shows office tasks such as cleaning trash, opening a box, and folding clothing. Some of the demonstrations were absent from the pre-training data. After post-training with less than one hour of task-specific data, the robot could repeat highly dexterous tool-use tasks without failure in the displayed examples. He presents this as a route toward commercial deployment, where the same physical task must be repeated many times. A model that only performs once for a video is insufficient. The broad pre-training gives the model physical knowledge before the task-specific examples arrive, so the team does not need to collect a large dataset from scratch for every new task.

06:10

An 80 to 90 percent success rate fails when a task must repeat many times

Ma says generalist manipulation models often reach an 80 to 90% success rate in deployment. That sounds strong in a single demonstration, but the chance of completing the same task ten times in a row is less than 0.1% at those rates. Specialist systems can perform one task reliably, yet their pipelines do not transfer quickly to arbitrary new tasks. Dyna is trying to combine the breadth of a generalist model with the reliability needed for commercial work. Ma frames this as the central engineering problem: the robot must handle many tasks while recovering from the mistakes that broad models inevitably make.

07:40

Napkin folding exposes small physical errors and strict commercial quality standards

Dyna chose restaurant napkin folding as an early case study because it is repetitive, commercially useful, and harder than it looks. The robot must pull exactly one napkin from a stack, fold it, and place it in a bin. Parallel-jaw grippers may pull out extra napkins, which creates a recovery problem. Restaurants also have an explicit quality threshold. Ma compares an accepted grade-five fold with a grade-three fold that fails acceptance because the second fold seam is slightly misplaced. Dyna-1 reached a reported 99.4% success rate over a 24-hour span. Ma says the result came from four distinct 24-hour trials, with changing light in the environment.

11:39

Reward models turn robot mistakes into targeted training examples

Dyna trained a reward model that watches a robot video and scores progress toward completing a long-horizon task. A successful run has progress that rises roughly monotonically from zero to one. When the robot starts making a mistake, the score shows a dip or other non-monotonic behavior. That signal lets researchers find difficult cases without having people watch every run continuously. Operators can collect the specific recovery data associated with the failure, fine-tune the robot, and repeat the process. Ma calls this scalable supervision and describes it as a human-in-the-loop active-learning cycle. The reward model does not replace human input. It directs human effort toward the cases where new data is most useful.

14:08

Recovery data helps the model handle errors it has never seen exactly

Napkin folding involves deformable objects, so a napkin can reach almost infinitely many configurations. Dyna cannot collect an example for every possible error. After repeated active-learning cycles, Ma says the model began to generalize recovery behavior to new mistakes. In one example, the robot pulled out more than one napkin and separated them before continuing. In another, it pulled over the entire stack, made a large mess, and still recovered through pulling and stretching actions. These cases matter because standard post-training often leaves the robot stuck after its first mistake. The general pre-training also contributes recovery behavior learned from other tasks, according to Ma's later answer.

16:28

Dyna tested site generalization with T-shirt folding at CoRL 2025

The napkin and laundromat demonstrations used data collected at the customer site. Ma then describes a separate recipe intended to let robots work in unseen environments without additional data. At CoRL 2025 in Korea, Dyna brought a robot from the United States to an exhibition booth and had it fold T-shirts. Attendees deliberately interfered, including covering the camera with a T-shirt. The robot continued folding for three days. Ma presents this as evidence that the model can transfer a task to a new site rather than requiring a fresh collection and fine-tuning cycle for every deployment. He also mentions a Red Bull partnership in which robots opened cans at changing-light events.

"So whenever a model does make a mistake, we as researchers or operators can immediately know the kind of video case that the model is struggling on and then do very targeted data collection and error recovery data for the model."13:25
Who should watch
  • You are building robot foundation models and need a training loop for failures that appear after deployment.
  • Your robot performs well in demos but stalls when a mistake pushes it outside the training distribution.
  • You are evaluating whether broad pre-training can reduce the site-specific data needed for commercial robot deployments.