Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics

World-action models represent a paradigm shift in robotics by integrating future state prediction with physical action generation. Dyna-2 utilizes a diffusion transformer architecture trained on over one million hours of data to enable precise, dexterous motion in real-world environments. This approach emphasizes the importance of high-fidelity, egocentric data and hardware-software co-design, allowing robots to perform complex, multi-step tasks like laundry folding and food preparation with human-like adaptability. By controlling the entire stack—from actuators to inference engines—the system achieves low-latency performance and consistent execution across diverse sites. Future advancements hinge on scaling synthetic data generation and incorporating multi-modal sensory inputs, such as tactile and auditory feedback, to move beyond current vision-centric limitations and achieve truly general-purpose robotic autonomy.