Zero-shot stitching in Reinforcement Learning using Relative Representations

Antonio Pio Ricciardi, Valentino Maiorca, Luca Moschella, Emanuele Rodolà

Keywords: visual reinforcement learning, reinforcement learning, latent space translation, relative representation, zero-shot, stitching, latent communication

TL;DR: Extended abstract: we perform zero-shot stitching between policies and encoders trained on variations of the Car Racing environment

Abstract:

In this paper we investigate the use of a recent method called "relative represen-
tations" to enable zero-shot model stitching in visual RL between encoders and
policies trained on the CarRacing environment, which does not require additional
training. Our experiments show that the relative representation framework can
be effectively applied to the RL realm to obtain compositionality and therefore
zero-shot stitching across agents with multiple variation factors: i) random seed for
the training; ii) environment style (background color); iii) training algorithm used
(PPO and DDQN)