English

Offline Reinforcement Learning at Multiple Frequencies

Machine Learning 2022-07-27 v1 Artificial Intelligence Robotics

Abstract

Leveraging many sources of offline robot data requires grappling with the heterogeneity of such data. In this paper, we focus on one particular aspect of heterogeneity: learning from offline data collected at different control frequencies. Across labs, the discretization of controllers, sampling rates of sensors, and demands of a task of interest may differ, giving rise to a mixture of frequencies in an aggregated dataset. We study how well offline reinforcement learning (RL) algorithms can accommodate data with a mixture of frequencies during training. We observe that the QQ-value propagates at different rates for different discretizations, leading to a number of learning challenges for off-the-shelf offline RL. We present a simple yet effective solution that enforces consistency in the rate of QQ-value updates to stabilize learning. By scaling the value of NN in NN-step returns with the discretization size, we effectively balance QQ-value propagation, leading to more stable convergence. On three simulated robotic control problems, we empirically find that this simple approach outperforms na\"ive mixing by 50% on average.

Keywords

Cite

@article{arxiv.2207.13082,
  title  = {Offline Reinforcement Learning at Multiple Frequencies},
  author = {Kaylee Burns and Tianhe Yu and Chelsea Finn and Karol Hausman},
  journal= {arXiv preprint arXiv:2207.13082},
  year   = {2022}
}

Comments

Project website: https://sites.google.com/stanford.edu/adaptive-nstep-returns/

R2 v1 2026-06-25T01:15:01.466Z