Model-Based Reinforcement Learning with Multi-Task Offline Pretraining

TL;DR

Vid2Act addresses the challenge of transferring knowledge from offline multi-task data to a new online visual control task by leveraging a mixture, action-conditioned world model as both a dynamics predictor and a task relevance evaluator. It introduces domain-selective dynamics distillation to transfer useful dynamics from multiple source tasks and a domain-selective behavior transfer via a frozen action-replay model to guide the target policy with relevant source actions. The approach demonstrates strong improvements over state-of-the-art model-based RL baselines on Meta-World, DM Control, and CARLA, including positive transfer even when source data is partially unrelated. While offering practical gains for data-efficient RL, Vid2Act incurs substantial offline pretraining time due to jointly learning dynamics across multiple tasks within the mixture world model.

Abstract

Pretraining reinforcement learning (RL) models on offline datasets is a promising way to improve their training efficiency in online tasks, but challenging due to the inherent mismatch in dynamics and behaviors across various tasks. We present a model-based RL method that learns to transfer potentially useful dynamics and action demonstrations from offline data to a novel task. The main idea is to use the world models not only as simulators for behavior learning but also as tools to measure the task relevance for both dynamics representation transfer and policy transfer. We build a time-varying, domain-selective distillation loss to generate a set of offline-to-online similarity weights. These weights serve two purposes: (i) adaptively transferring the task-agnostic knowledge of physical dynamics to facilitate world model training, and (ii) learning to replay relevant source actions to guide the target policy. We demonstrate the advantages of our approach compared with the state-of-the-art methods in Meta-World and DeepMind Control Suite.

Figures (10)

• Figure 1: We aim to build an offline-to-online transfer RL agent for visual control problems, which is challenging due to the discrepancies between the target task and the source tasks from which the offline datasets are collected. The key idea of our approach is to leverage the world models to enable positive knowledge transfer through domain-selective dynamics distillation and behavior guidance.
• Figure 2: Left: We employ multiple offline domains to train a mixture world model ($F_\phi$), whose parameters are frozen during the subsequent transfer learning process. Right: In the target domain, we use $F_\phi$ as the teacher model and dynamically distill prior knowledge from it with a set of domain-similarity weights $\mathcal{W}$. These weights are further used to reproduce the most relevant source actions to guide the target policy.
• Figure 3: Showcases of selected towns in CARLA environment.
• Figure 4: Performance comparison with the state-of-the-art methods on Meta-World as measured on the success rate. Vid2Act outperforms the compared models.
• Figure 5: Performance comparison on two tasks from DeepMind Control Suite as measured on the episode rewards. Our Vid2Act with dynamic knowledge distillation achieves significant improvements compared with existing model-based RL approaches.