Dyna Robotics Achieves 99.4% Success in Napkin Folding

Jason Ma, co-founder and CTO of Dyna Robotics, leads a company established in September 2024, which has already secured $120 million in Series A funding.

The company's core mission is to develop high-performance, robust foundation model autonomy that can be effectively deployed in real-world applications, employing a 'research and deployment flywheel' to connect lab innovations with practical commercial utility.

Dyna Robotics currently operates more than five deployment sites to test various workflows and gather high-quality data.

These real-world deployments are crucial for identifying specific weaknesses in existing models and hardware, allowing the company to move beyond simple YouTube demonstrations towards scalable solutions that have a tangible impact on real-world environments and millions of people's lives.

The 'data pyramid' strategy employed by Dyna Robotics integrates off-robot data, diverse on-robot task data, and high-quality deployment data.

While off-robot data, including human-captured videos and public datasets, lacks precision for robotic actions, it is combined with on-robot data from industrial, household, and hospitality settings, and critical deployment data, to effectively close the distribution gap between laboratory facilities and real-world applications; the training pipeline currently processes over 200,000 hours of data.

The model architecture at Dyna Robotics uses a dual-layer approach, combining a high-level reasoning model with a low-level world action model.

This allows the robot to maintain semantic context while executing high-frequency, precise actions, which is essential for recovering from physical errors during complex task execution.

With less than an hour of task-specific data, Dyna Robotics' models perform dexterous work, including tasks outside the original training set.

This rapid deployment capability, coupled with the necessity for high reliability, is crucial for transitioning from basic demonstrations to commercially viable, failure-free operations in human environments.

Current foundation models typically achieve only an 80-90% success rate in real-world deployments.

This success rate translates to less than a 0.1% chance of completing ten consecutive tasks without failure, highlighting a significant reliability gap.

While specialized robot pipelines offer high reliability for singular tasks, they lack scalability; Dyna Robotics aims to merge general-purpose capabilities with the commercial-grade reliability necessary for practical applications.

Napkin folding was selected as a critical case study due to its commercial ubiquity and inherent complexity.

The Dyna-1 model achieved a 99.4% success rate over a continuous 24-hour period, demonstrating its ability to adapt to shifting environmental conditions, such as changes in lighting.

Failure recovery was a crucial aspect, particularly when the grippers failed to pick exactly one napkin, which the system successfully managed.

Unlike lab-based experiments, commercial environments demand precise and well-defined success criteria for robotic tasks.

For napkin folding, a 'grade five fold' represents the minimum quality acceptable in a restaurant setting, often differing from a failed 'grade three fold' by a marginal one-inch seam deviation.

Standard pre-training and post-training methods alone typically result in only an 80% success rate for robotic tasks.

Models trained conventionally often struggle to remain within distribution once an error occurs, leading to situations where they get stuck and are unable to recover from initial mistakes.

Reward models enable 'scalable supervision' by automating the detection of failures, allowing robots to identify when they make mistakes.

Human operators then engage in targeted data collection for specific error cases, addressing the model's struggles.

This active learning cycle allows the model to iterate and significantly enhance its robustness through fine-tuning.

Napkin folding, as a deformable object manipulation task, involves an almost infinite number of possible states, making exhaustive data collection for every error case impractical.

Through iterative active learning, the robot developed the ability to generalize and recover from novel mistakes, demonstrating significant progress; during a 24-hour trial, the robot successfully recovered even after accidentally tipping over an entire stack of napkins, achieving a 99.4% success rate.

Dyna Robotics has successfully deployed napkin-folding robots in restaurants across the United States, managing tasks for customers in back offices.

The company has also expanded its capabilities to towel folding at a laundromat in Sacramento, demonstrating that the model can be fine-tuned for various site-specific needs.

Dyna Robotics developed a data recipe to generalize robot performance to new environments without requiring additional training, thus maintaining task performance beyond pre-collected site-specific data.

The robot was showcased at CoRL 2025 in Korea, where it folded t-shirts for three days in a public exhibition, demonstrating its durability against real-world interferences like camera obstructions.

The recipe allowed for immediate deployment, enabling the robot to start folding various t-shirts as soon as it was placed.

Dyna Robotics partnered with Red Bull to deploy robots for opening drink cans at music festivals, showcasing their performance under challenging and fluctuating lighting conditions.

This initiative further reinforced the company's commitment to building general-purpose models capable of high-performance commercial execution in diverse and unpredictable environments.

Dyna Robotics is currently focused on developing its own proprietary hardware stack rather than offering developer kits at this time, though such kits are on their future roadmap.

Dyna Robotics' long-term vision is to develop a general-purpose robot model and hardware platform deployable anywhere for anyone, including consumers.

However, the current strategy prioritizes enterprise customers, as commercial settings serve as essential testing grounds for iterating and deploying robotic solutions, especially given the higher polish, lower error tolerance, and strict privacy compliance required for consumer products.

Dyna Robotics aims to create teachable robots capable of interpreting verbal human instructions, leveraging existing speech-to-text models to quickly translate human intent into text.

This translation then feeds into their reasoning and world models for action execution, though the primary bottleneck remains the foundation model's ability to effectively perform low-level robotic manipulation tasks.

A generalist foundation model provides a crucial base of physical and semantic understanding, enhancing robustness by allowing the model to apply error-recovery behaviors learned from diverse tasks.

Specializing a model on a single task from the outset risks missing out on valuable cross-task physical knowledge that contributes to greater overall robustness.

Dyna Robotics utilizes a comprehensive approach involving pre-training, post-training, and active learning to achieve commercial-grade reliability across a wide range of applications.