PriPG-RL: Privileged Planner-Guided Reinforcement Learning for Partially Observable Systems with Anytime-Feasible MPC

TL;DR

This paper introduces PriPG-RL, a framework that leverages a privileged planner during training to enhance reinforcement learning in partially observable systems, validated through simulation and real-world deployment.

Contribution

It presents a novel training framework combining a privileged planner with RL, including a new MPC algorithm and a distillation method for improved policy learning under partial observability.

Findings

Enhanced sample efficiency and policy performance demonstrated in simulations.

Successful deployment on a quadruped robot navigating complex environments.

Theoretical analysis supporting the proposed framework.

Abstract

This paper addresses the problem of training a reinforcement learning (RL) policy under partial observability by exploiting a privileged, anytime-feasible planner agent available exclusively during training. We formalize this as a Partially Observable Markov Decision Process (POMDP) in which a planner agent with access to an approximate dynamical model and privileged state information guides a learning agent that observes only a lossy projection of the true state. To realize this framework, we introduce an anytime-feasible Model Predictive Control (MPC) algorithm that serves as the planner agent. For the learning agent, we propose Planner-to-Policy Soft Actor-Critic (P2P-SAC), a method that distills the planner agent's privileged knowledge to mitigate partial observability and thereby improve both sample efficiency and final policy performance. We support this framework with rigorous theoretical analysis. Finally, we validate our approach in simulation using NVIDIA Isaac Lab and successfully deploy it on a real-world Unitree Go2 quadruped navigating complex, obstacle-rich environments.