Robot Demos Are Easy. Reliability Is Hard — Jason Ma, Dyna Robotics [26:41]

**TL;DR:** Dyna Robotics addresses the "reliability gap" in robotics by combining a large-scale data pyramid, a specialized reward model for active learning, and iterative deployment to achieve near-perfect (99.4%) task performance, moving beyond impressive demos to commercial-grade autonomy. ## The Mission: Building Robust Foundation Models for Robotics Dyna Robotics was founded in September 2024 with the mission to build highly robust foundation models for autonomous robots in the real world. The company, which raised approximately $120 million in its Series A, aims to create a single platform that can perform a variety of economically valuable physical tasks. Their core thesis is that achieving true commercial viability requires merging cutting-edge research with large-scale deployment. ## The Research & Deployment Flywheel To break through the challenges of robotic deployment, Dyna operates on a "research & deployment flywheel." This involves: 1. **Internal R&D & Hardware Innovation:** Guides which commercial tasks and workflows to target. 2. **Active Deployment:** With over five active sites, the company collects high-quality deployment data. 3. **Product & Data Feedback:** Deployment reveals model and hardware shortcomings, focusing research on the most critical, real-world problems that separate YouTube demos from impactful applications. ## Training General-Purpose Robot Models ### The Data Pyramid Due to the scarcity of robot data, Dyna trains its models using a multi-tiered data pyramid: * **Non-robot Data:** Includes human data from wear cameras, public datasets, and simulation. * **Robot Data:** Collected across diverse tasks and environments (industrial, home, laundry, hotels). * **Deployment Data:** High-quality data from real-world use to bridge the simulation-to-real gap. The training pipeline has accumulated over **200,000+ hours** of data. ### Model Architecture & Capabilities The model architecture consists of a high-level reasoning model and a lower-level "world action model" for fine-grained, dexterous control. This design is essential for physical tasks requiring both semantic understanding and precise interaction to recover from errors. With strong pre-training, the model can be fine-tuned for new tasks with **less than one hour** of task-specific data. However, these general models typically achieve only **80-90% reliability** in practice, making them unsuitable for continuous real-world operation. ## The Core Problem: Achieving 100% Robustness Dyna's central challenge was to transform a general-purpose pre-trained and fine-tuned model into one that achieves **near-100% robustness** for a specific, valuable task. They selected **napkin folding** for restaurants—a repetitive, mundane task with clear commercial demand and well-defined success criteria (e.g., a "five-tier fold" vs. a sloppy "three-tier fold"). ### The Dyna-1 Model: A Case Study in Reliability The result was the **Dyna-1 model**, a fine-tuned version of their foundation model dedicated to folding napkins. In a 24-hour endurance test, it achieved a **99.4% success rate**. This test was not a one-off; it was repeated successfully four times. The video timelapse showed the model operating robustly even as environmental lighting changed from night to day. **Task Complexity:** The robot must precisely pick a single napkin from a stack using a parallel gripper, fold it, and place it in a bin. Failures often stem from gripper imprecision, requiring the model to recover from errors like picking up multiple napkins. ## How They Achieved Near-Perfect Reliability ### Overcoming the Limits of Standard Fine-Tuning Initially, using standard "pre-train + fine-tune" methods yielded only **80% success**. Once the model made an error, it would drift out of its training distribution and fail to recover. ### The Solution: A Reward Model for Scalable Oversight To solve this, Dyna developed a specialized **reward model** that can assess task progress by analyzing robot videos. This model outputs a score from 0 (task start) to 1 (completion). Crucially, the score **dips** when the robot makes a mistake, providing a signal for failure detection. ### The Active Learning Loop This reward model enables "scalable oversight" and a powerful active learning cycle: 1. **Autonomous Execution:** The robot attempts to fold napkins while the reward model runs in the background. 2. **Automatic Error Detection:** When the reward model's score indicates an error, the case is flagged. 3. **Targeted Data Collection & Training:** Researchers or operators collect specific data on the error and train the model for that failure mode. 4. **Iterative Improvement:** The model is re-fine-tuned with this new data, strengthening its error recovery capabilities. After several iterations of this loop, the model becomes exceptionally robust, able to recover from a wide variety of unexpected errors, such as accidentally grabbing multiple napkins or even knocking over the entire stack. This continuous error-correction ability is what allowed Dyna-1 to achieve **99.4% reliability** over 24 hours, demonstrating a clear path to commercial-grade performance. **Source:** [Robot Demos Are Easy. Reliability Is Hard — Jason Ma, Dyna Robotics](https://www.youtube.com/watch?v=Sjfz1TqxzEs)