A-NC: Adaptive Neural Control with implicit online inference of privileged parameters

Rapid changes in the environment and robot parameters pose significant challenges for control systems, particularly when key parameters are not directly measurable. In this paper, we introduce a novel approach using classical Recurrent Neural Network (RNN) controllers to dynamically adapt policies in response to these changes. We propose strategies for data collection and processing that enable the successful training of efficient Gated Recurrent Unit (GRU) nonlinear controllers capable of adapting to changing parameters. We demonstrate this approach using a simulated and a physical cartpole robot. The RNNs are trained through supervised learning on data generated in simulation using Nonlinear Model Predictive Control (NMPC). We vary the cartpole’s angle sensor offset or pole length jointly with pole mass, none of which are directly measurable by the robot. Our results show how the RNN controller adjusts its policy based on past trajectories, leading to control that mimics the NMPC, outperforming Domain Randomization (DR) technique applied to feedforward neural networks. Unlike NMPC, which relies on explicit knowledge of environment parameters, the RNN implicitly estimates these parameters from past trajectories, allowing it to adapt its control policy dynamically. It also outperforms NMPC control performance when the parameters relevant for NMPC are not known.