DistRL: An Asynchronous Distributed Reinforcement Learning Framework for On-Device Control Agents

Taiyi Wang , Zhihao Wu , Jianheng Liu , Jianye Hao , Jun Wang , Kun Shao

🏛 Institutions: University of Cambridge , Powersense Technology Limited , Huawei Noah's Ark Lab , UCL , Tianjin University
📅 Date: October 18, 2024
📑 Publisher: ICLR 2025 (Poster)
💻 Env: Mobile
🔑 Keywords: framework reinforcement learning distributed RL fine-tuning centralized training decentralized data acquisition A-RIDE DistRL

TLDR

DistRL is a distributed RL fine-tuning framework for mobile control agents that separates centralized training from decentralized data collection across worker devices. It is paired with the A-RIDE off-policy RL algorithm, and the paper reports 3x higher training efficiency, 2.4x faster data collection, and a 20% relative success-rate gain on open Android control tasks.

Open paper Report issue