Adapting Neural Models with Sequential Monte Carlo Dropout

Neural models and policies are now ubiquitous in modern robotics. The prevailing approach to training these follows a two stage process - a large, comprehensive collection of data (often state and action pairs) is used to train a suitable model or policy, which is then frozen and deployed. Unfortunately, this results in models that are unable to adapt to changes in the environment, which is a particular concern in robotics. For example, it would be preferable for a robot dynamics model to handle context dependent kinematic or dynamic properties, or a collaborative robot relying on predictions of human behaviour to adapt to different human abilities or preferences. Many existing adaptive control techniques [1] attempting to tackle this problem rely on carefully considered parametric models, but these may lack the requisite capacity for prediction that is typically associated with neural models. In contrast, meta-learning and adaptive neural control approaches addressing this problem are often quite cumbersome to train and implement. This paper introduces a simple and effective approach to achieve adaptation for neural network models.