high performance computing on graphics processing units: hgpu.org

Data-efficient learning in continuous state-action spaces using high-dimensional observations remains an elusive challenge in developing fully autonomous systems. An instance of this challenge is the pixels to torques problem, which identifies key elements of an autonomous agent: autonomous thinking and decision making using sensor measurements only, learning from mistakes, and applying past experiences to novel situations. In this research, we introduce a deep dynamical convolutional model, able to learn complex non-linear dynamics and do long-term predictions. Compared to state-of-the-art reinforcement learning methods for continuous state and action space problems, our approach is solid and efficient as it is model-based, is scalable to high-dimensional state spaces, learns quickly, and is a major step towards fully autonomous learning from pixels to torques.