Robot Demos Are Easy. Reliability Is Hard — Jason Ma, Dyna Robotics

Jason Ma, co-founder and CTO of Dyna Robotics, discusses the company’s approach to developing high-performance and robust generalist robotics policies for real-world commercial deployment. Dyna Robotics aims to build foundation model autonomy that enables a single platform to perform numerous economically useful physical tasks.

The Research and Deployment Flywheel

Founded in September 2024 (likely a transcription error for an earlier year given the context of Series A funding and deployment timelines), Dyna Robotics operates on a research and deployment flywheel model. Frontier research and in-house hardware development inform real-world task commercialization. Active deployment at over five sites serves multiple purposes:

• Gathering high-quality deployment data.
• Identifying model and hardware limitations.
• Sharpening research focus on critical real-world robotics problems, moving beyond mere demos to impact.

Robot Foundation Models and Data Pyramid

Dyna’s robot foundation models for manipulation are trained on a pre-training data pyramid totaling over 200,000 hours of data:

1. Diverse Off-Robot Data: Collected from humans wearing cameras, public datasets, and simulation data, addressing data scarcity.
2. Diverse On-Robot Data: Task-specific data collected across various scenarios (industrial, household, laundromat, hotels) on physical robots.
3. High-Quality Deployment Data: Collected from live deployments to close the train-test distribution gap, as lab environments differ significantly from real-world settings.

The model architecture combines a high-level reasoning model with a low-level world action model. This dual approach is crucial for physical tasks, requiring both semantic understanding and fine-grained physical interaction to recover from mistakes and execute precise actions. This comprehensive pre-training allows models to be rapidly fine-tuned for diverse tasks, often requiring less than an hour of task-specific data for new tasks not present in the pre-training set.

The Challenge: Reliability for Commercial Deployment

Ma emphasizes that while generalist models can produce impressive demos with 80-90% success rates, this level of reliability is insufficient for commercial use. A model with an 80% success rate has less than a 0.1% chance of completing the same task 10 times in a row. Traditional specialized robots achieve high reliability for single tasks but lack scalability. Dyna’s mission is to resolve this dichotomy by creating general-purpose models that are both versatile and reliable enough for commercial deployments.

Achieving Near-100% Robustness: The Reward Model Approach

Dyna’s breakthrough for achieving extremely high robustness involves reward models and a human-in-the-loop active learning cycle.

Case Study: Restaurant Napkin Folding (Dyna-1)

• Dyna-1, a generalist foundation model fine-tuned for napkin folding, achieved a 99.4% success rate over a 24-hour span. This includes continuous operation through changing lighting conditions (e.g., from night to day).
• Task Complexity: Napkin folding requires precise picking of a single napkin from a stack, accurate folding to meet strict quality criteria (e.g., Grade 5 fold vs. Grade 3), and robust recovery from common errors like picking multiple napkins. Deformable object manipulation (like napkins) presents an almost infinite number of possible states, making exhaustive error data collection impossible.
• Reward Models: Dyna developed models that can analyze robot video and accurately score progress toward task completion. When the robot makes a mistake, the progress estimation shows a non-monotonic dip, indicating an error.
• Scalable Supervision and Active Learning: This allows operators to automatically identify failure cases. Instead of constant manual oversight, the reward model pinpoints when and how the robot struggles. This enables targeted data collection for error recovery and subsequent fine-tuning. This human-in-the-loop active learning process iteratively improves the model’s robustness and ability to generalize to new error recovery methods.
• Error Recovery Highlights: The Dyna-1 demonstrated impressive recovery behaviors, such as separating multiple picked napkins or even autonomously re-orienting and continuing the task after accidentally pulling over an entire stack of napkins.

Generalization and Real-World Deployments

Initially, deployments required data collection at the customer site. However, Dyna has developed a data recipe enabling models to generalize to new, unseen environments without additional data while maintaining performance.

• Commercial Deployments: Robots are successfully folding napkins in restaurants and towels in a Sacramento laundromat.
• CoRL 2025 Demo (Korea): A robot folded T-shirts for three days straight at the conference, enduring interference from attendees and demonstrating robust operation in an unfamiliar environment without site-specific data.
• Red Bull Events: Robots are deployed to open Red Bull cans at live events, handling dynamic lighting and diverse environments.

Future Outlook and Q&A

Dyna Robotics is focused on enterprise use cases for now, as they provide ideal testing grounds for co-developing AI and hardware. While developer kits and direct-to-consumer products are long-term considerations, the current emphasis is on refining the core technology. Regarding human-robot interaction, Ma foresees integrating speech-to-text models with their reasoning and world models to enable voice commands, though low-level manipulation capabilities remain a key bottleneck. The generalist foundation model approach is crucial for building robust error recovery skills by interpolating knowledge from a vast array of tasks.

• Tags:
• Foundation Models
• Robotics
• Human-in-the-Loop
• Active Learning
• Real-World AI
• Manipulation
• Reward Models
• Error Recovery
• Commercial AI
• Robot Deployment
• Dyna Robotics