Efficiently Learning Small Policies for Locomotion and Manipulation
Shashank Hegde Gaurav S. Sukhatme
University of Southern California
Abstract
Neural control of memory-constrained, agile robots requires small, yet highly performant models. We leverage graph hyper networks to learn graph hyper policies trained with off-policy reinforcement learning resulting in networks that are two orders of magnitude smaller than commonly used networks yet encode policies comparable to those encoded by much larger networks trained on the same task. We show that our method can be appended to any off-policy reinforcement learning algorithm, without any change in hyperparameters, by showing results across locomotion and manipulation tasks. Further, we obtain an array of working policies, with differing numbers of parameters, allowing us to pick an optimal network for the memory constraints of a system. Training multiple policies with our method is as sample efficient as training a single policy. Finally, we provide a method to select the best architecture, given a constraint on the number of parameters.
Video Summary
Rollouts
For each task we find the smallest network architecture that at least achieves 90% peak performance
Hopper [8,4,4,4] - 187 parameters
HalfCheetah [64,4,4,32] - 1790 parameters
Walker2D [16,16,4] - 658 parameters
Humanoid [8,32,8] - 3721 parameters
Ant [16,32,32,4] - 3564 parameters
FetchReach [4,8] - 132 parameters
FetchPush [32,32,8,16] - 2460 parameters
FetchSlide [64,64,32,16] - 8692 parameters
FetchPickAndPlace [16,16,8] - 824 parameters
Method Overview
GHP control policy weight estimation: The GHP can estimate the desired weights for a given Control policy network architecture. This way we can estimate the optimal control policy weights for multiple architectures.
Off-Policy RL using the GHP: For an existing Actor Critic, Off-Policy Reinforcement Learning algorithm, we replace the vanilla actor policy with a Graph Hyper Policy (GHP).
Using DLM for best architecture: Among architectures trained with the GHP, we can predict the best performing architecture by maximizing the DLM
Citation
@inproceedings{hegde2023efficiently,
title={Efficiently Learning Small Policies for Locomotion and Manipulation},
author={Hegde, Shashank and Sukhatme, Gaurav S},
booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
pages={5909--5915},
year={2023},
organization={IEEE}
}