Predictive Neural Network Models For Autonomous Vehicle Driving on Multiple Terrains

This work considers the problem of predicting the dynamics of a small unmanned ground vehicle (UGV) navigating on unknown terrains. Predicting the dynamics over long horizons accurately helps avoid obstacles by making accurate control decisions early on. We propose a Recurrent Neural Network (RNN) to predict the position and orientation of the vehicle driving at 8 meters per second (18 mph) up to a time horizon of 2 seconds. Our method is different from traditional dynamic models of a car in the sense that we do not explicitly encode the terrain dependent parameters into the model. In contrast, the RNN model can estimate both the type of the terrain and the terrain parameters by processing through a few steps of sensor data such as GPS, IMU and wheel encoders. We tested the network on a 1/5th scale RC car driving on three terrain surfaces namely grass concrete and sand which have vastly different dynamic behaviors. Our network is able to make position predictions with an accuracy of 0.5 m and orientation of the car with an accuracy of 7 degrees in the next two seconds. In addition, the network is also able to classify the terrain accurately 74% of the time (3 terrains considered) where random chance would predict at 33%. The dynamic model created in this work can be used in predictive control schemes such as Model Predictive Control (MPC) or to simply determine the safe driving conditions of a vehicle based on the terrain.