CLIP-Motion: Learning Reward Functions for Robotic Actions Using Consecutive Observations

This paper presents a novel method for learning reward functions for robotic motions by harnessing the power of a CLIP-based model. Traditional reward function design often hinges on manual feature engineering, which can struggle to generalize across an array of tasks. Our approach circumvents this challenge by capitalizing on CLIP's capability to process both state features and image inputs effectively. Given a pair of consecutive observations, our model excels in identifying the motion executed between them. We showcase results spanning various robotic activities, such as directing a gripper to a designated target and adjusting the position of a cube. Through experimental evaluations, we underline the proficiency of our method in precisely deducing motion and its promise to enhance reinforcement learning training in the realm of robotics. Reinforcement Learning (RL) distinguishes itself within the machine learning spectrum by enabling an agent to determine optimal decisions through interactions with an environment, while targeting maximal cumulative reward. The linchpin of this approach is the reward mechanism, which steers the agent's decision-making. Each reward, be it positive or negative, influences the agent's action choices. Essentially, the reward system is RL's guiding compass, directing the agent toward optimal actions and ensuring its adaptability in dynamic and unpredictable scenarios. Without it, the agent would drift aimlessly, unable to discern advantageous actions from detrimental ones.