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

## Executive summary

Dyna Robotics focuses on developing highly robust, generalist robotic policies for commercial deployment, arguing that high reliability is more critical than high performance in demos. The company utilizes a 'research and deployment flywheel' to build foundation models, achieving a 99.4% success rate in complex tasks like napkin folding over 24 hours. Key technical advancements include a 'pre-training data pyramid' (over 200,000 hours) and the use of reward models for scalable supervision, allowing the system to detect and recover from errors in long-horizon tasks.

## Key takeaways

- The Reliability Gap in Robotics: Achieving a high success rate (e.g., 99.4%) over extended periods (24 hours) is necessary for commercial viability, as standard models often stall at 80–90% success rates, making repeated tasks highly improbable.
- The Research and Deployment Flywheel: Dyna Robotics combines frontier research with active commercial deployments to gather high-quality data, which informs and sharpens the focus of their model development, ensuring the product solves real-world problems.
- Scalable Error Recovery via Reward Models: Instead of relying on manual oversight, the team developed reward models that score the robot's progress during complex tasks. Dips in this score signal a mistake, enabling targeted data collection and a human-in-the-loop active learning cycle for robust error recovery.
- Generalization Across Sites: The model architecture is designed to generalize, allowing deployment at new customer sites (e.g., a laundromat, Red Bull events) without requiring site-specific fine-tuning or additional data.

## Technical details

- Model Architecture: The system uses a combination of a high-level reasoning model and a low-level world action model. This pairing allows the robot to maintain semantic understanding while executing fine-grained, high-frequency physical actions.
- Data Collection Strategy (Pre-training Data Pyramid): To overcome the scarcity of robotics data, the model is trained using a pyramid approach combining three sources: 1) Diverse off-robot data (human-worn cameras, public datasets); 2) On-robot data (industrial, household, laundromat tasks); and 3) High-quality deployment data (closing the train/test distribution gap). The pipeline currently holds over 200,000 hours of data.
- Task Specificity and Generalization: The model can be rapidly fine-tuned for new tasks using very little data (less than 1 hour of task-specific data). Furthermore, error recovery behavior is often achieved by interpolating knowledge gathered from thousands of different tasks, rather than being limited to the specific task being performed.

## Practical implications

- The focus on robust, generalist models suggests that future build-engineering efforts should prioritize reliability and error recovery mechanisms over maximizing task throughput in controlled environments.
- The 'Research and Deployment Flywheel' model provides a blueprint for productizing AI robotics, emphasizing that commercial deployment data is as valuable as lab data.
- The use of reward models for active learning offers a scalable method for quality assurance in complex robotic systems, allowing engineers to pinpoint failure modes efficiently.

## Topics

Robotics, Artificial Intelligence, Foundation Models, Machine Learning, Active Learning, Build Engineering, Dyna Robotics, Dyna-1

Source: https://www.youtube.com/watch?v=Sjfz1TqxzEs