Skill Acquisition with Switching Latent PI Controllers

Juyan Zhang, Dana Kulic, Michael Burke

Keywords: Deep learning, Variational Inference, Skill learning, Controller

TL;DR: Learning skills in latent space with latent PI controller

Abstract: Recent work recasts neural network mixture heads as a library of latent-space feedback skills, with each skill behaving as a proportional (P) controller and achieving improved robustness compared to behavior cloning baselines. We extend this framework to proportional–integral (PI) control in latent space and evaluate our model on FetchPush and robot-trajectory tasks and compare it with P controller and other Behaviour cloning baselines. Results show that the integral path accumulates latent tracking error to cancel slowly varying disturbances, which empirically enhances robustness and performance at the cost of modestly lower sample efficiency in the lowest data setting. We also find that the introduction of the PI controller empirically bounds the deployment trajectory closer to the training trajectory than the P controller

Submission Number: 9