Adaptive solution to transfer learning of neural network controllers from earth to space environments

Abstract

AbstractCompliant manipulation has long been a major constraint for grappling in robotic manipulators. To adopt robotic manipulators in space for the prospect of capturing space junk and transforming them into salvageable assets for re‐use, robust adaptive manipulation would be key. We believe that a bio‐inspired approach could provide human‐like tactility required for robustness and adaptability in robotic manipulation. Given the similarity in form and dynamics between earth‐based and space‐based robotic manipulators, we first explored the transfer learning of neural network controllers as an avenue to address the challenges of limited computation resources onboard the spacecraft (space manipulator). We introduced a pre‐trained and learned feedforward neural network for modelling the control error a priori. While the results were encouraging, there are major limitations of neural networks' capability to ensuring the transfer learning of similar earth‐based dynamics to space‐based dynamics, given that the parameters of contrast are fairly straightforward. We have demonstrated these limitations by presenting a novel approach that is inspired by human motor control. We explored the adaptability through a practical problem of transferring a neuro‐controller from earth to space. With the results not as plausible as expected, an alternative adaptive controller has been learned to demonstrate a viable solution. The controller was trained entirely in simulation via rapid online adaptation of the robot's controller to the object's properties and environmental dynamics using only proprioception history. As a notable step, we have shown that appropriate models can be learned in this manner by training the control policy via reinforcement learning, which provides avenue for transferring the learned model from earth to space environments.