23 February 2024
In robotics, deep learning excels in tasks such as vision-guided navigation, grasp-planning, human–robot interaction, and even design. Despite this, its application in generating motor intelligence in physical systems remains limited. Deep reinforcement learning, in particular, has shown the potential to outperform traditional approaches in simulations. However, its transfer to physical applications has been primarily hampered by the prerequisite of pretraining in a simulated environment.
The central drawback of general-purpose deep learning lies in its sample inefficiency, stemming from the need to distill all aspects of a task from data. In response to these challenges, there is a rising trend in robotics to specifically incorporate geometric priors into data-driven methods to optimize learning efficiency. This approach proves especially advantageous for high-level tasks that need not engage with the system’s physics. Physics-informed neural networks (PINNs), infusing fundamental physics knowledge into their architecture and training, have found success in various fields outside robotics, from earth science to materials science.
In this work, EMERGE partners from Delft University of Technology and collaborators deal with the experimental application of physics-informed neural networks to the modeling and control of complex robotic systems. Achieving this goal requires extending physics-informed neural networks to handle nonconservative effects.
These learned models are proposed to combine with model-based controllers originally developed with first-principle models in mind. By combining standard and new techniques, precise control performance can be achieved while proving theoretical stability bounds.
Read the paper in the link below.
More Information
