Physics-based learning for smarter motion controllers

点击翻译按钮获取中文摘要

Autonomous platforms are mostly limited by their ability to handle changing or unexpected situations. The proposed research aims at making them more autonomous and versatile, by investigating ways to include learning in the low-level motion controllers so that their “motor skills” can improve over time and adapt to new situations.State-of-the-art robots (self-driving cars, DARPA challenge humanoids, etc) typically use model-based control approaches which have been successful for controlling the motion of complex dynamic systems. The usual architecture consist of feedback laws and trajectory generation schemes relying heavily on dynamic models. However, with this type of controllers, robots don't improve their skills over time and usually struggle in environment that are complex and hard to model accurately. On the other hand, machine learning algorithms had a recent breakthrough with deep learning. However, training neural networks require huge databases of “questions and answers”, and is not suitable directly to robotics mobility problems. As stated by Sergey Levine, a leading researcher in the field: "Neural networks have made great strides in allowing us to build computer programs that can process images, speech, text, and even draw pictures. However, introducing actions and control adds considerable new challenges, [...] If we can bring the power of large-scale machine learning to robotic control, perhaps we will come one step closer to solving fundamental problems in robotics and automation."The proposed research aims at developing an hybrid control architecture to leverage the best of both worlds (model-based and learning), which could be the key for a breakthrough in terms of smarter motion controllers. More specifically, the short-term (5-years) objectives are to: 1) Propose and evaluate novel control architectures to include learning in the low-level feedback loops. For instance, event-based reflex reactions overriding a baseline model-based controller. 2) Explore the use of physics-based features for accelerating the learning and universalizing the knowledge. For instance, correlating appropriate maneuvers to dimensionless numbers (friction coefficient, Froude number, etc.) instead of platform-specific sensor stimulus. 3) Evaluate experimentally the developed control algorithms, with 1/10th scale autonomous cars, with the goal of learning to manage uncertain and challenging terrain conditions. Ultimately, the proposed research aims to develop enabling technologies in terms of robotics mobility that will contribute to bring automation in new fields such as transportation (self-driving cars), mining, agriculture, inspection, surveillance and even home use (autonomous vacuum cleaners and lawn mowers).