Achieving 99.4% Reliability in Generalist Robotics

The Reliability Gap in Robotics

Most generalist robot models currently plateau at 80–90% success rates, which is insufficient for commercial viability. At these rates, the probability of completing a sequence of ten tasks successfully drops below 0.1%. Dyna Robotics addresses this by focusing on "research and deployment flywheels," where real-world commercial deployments provide the high-quality, edge-case data necessary to move from demo-quality to production-grade performance.

The Data Pyramid and Architecture

To build robust models, Dyna utilizes a "pre-training data pyramid" consisting of three layers: diverse off-robot data (human-captured, public datasets, simulation), on-robot task data across varied industrial and household environments, and high-quality deployment data. Their architecture pairs a high-level reasoning model with a low-level world action model. This dual-layer approach allows the system to maintain semantic understanding while executing high-frequency, dextrous physical movements.

Reward-Driven Active Learning

To push past the 80% success ceiling, Dyna employs reward models that monitor robot progress in real-time. By scoring the robot's performance, the system identifies non-monotonic progress—a clear signal of failure. This enables a human-in-the-loop active learning cycle: when the reward model detects a dip, researchers collect targeted recovery data for that specific failure mode. This process was instrumental in achieving a 99.4% success rate in napkin folding over 24 hours, including autonomous recovery from catastrophic failures like toppling a stack.

Generalization Without Site-Specific Data

While early deployments relied on site-specific fine-tuning, Dyna has shifted toward a recipe that allows for zero-shot deployment. By training on a sufficiently diverse set of tasks and environments, the model can be deployed to new, unseen locations—such as the CoRL conference in Korea—and perform tasks like T-shirt folding for days without additional training, proving that the model has learned the underlying physics of the task rather than just memorizing a specific environment.