Practical Issues of Action-Conditioned Next Image Prediction

Abstract

The problem of action-conditioned image prediction in robotics is to predict the expected next frame given the current camera frame the robot observes and the action it selects. We provide the first comparison of two recent popular models, Convolutional Dynamic Neural Advection (CDNA) (6) and a feedforward model (15), especially for image prediction on cars. Our major finding is that action tiling encoding is the most important factor leading to the remarkable performance of the CDNA model. We present a light-weight model by action tiling encoding which has a single-decoder feedforward architecture same as (15). On a real driving dataset, the CDNA model achieves 0.3986 ×10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> MSE and 0.9836 Structure SIMilarity (SSIM) with a network size of about 12.6 million parameters. With a small network of fewer than 1 million parameters, our new model achieves a comparable performance to CDNA at 0.3613×10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> MSE and 0.9633 SSIM. Our model requires less memory, is more computationally efficient and more advantageous to be used inside self-driving vehicles.