Robot Modeling With Autoregressive Transformers

Many applications require accurate models of the dynamic behaviour of robots, such as realistic simulation, state prediction, and control. Traditionally, the modeling of these systems is based on analytical solutions that are complicated to develop. This is why in the state of the art some solutions are based on neural networks. We present a novel transformer-based architecture, called AR- Transformer, designed to estimate the model of complex nonlinear robotic systems. Our goal is to predict the behavior of the robot knowing only its initial state at any time step for any prediction horizon, knowing the external inputs or the control inputs. The AR- Transformer has an autoregressive approach, which allows the architecture to iterate and estimate states up to the time step for which the control inputs are known. This autoregressive approach works correctly with large prediction horizons thanks to training the model without teacher forcing. We have positively validated the architecture with a public dataset of real experiments of a 3 DOF robotic arm.