How to Train PointGoal Navigation Agents on a (Sample and Compute) Budget

TL;DR

This paper investigates training PointGoal navigation agents efficiently within limited sample and compute budgets, identifying key design choices that significantly improve performance across popular benchmarks.

Contribution

It provides a comprehensive analysis of training strategies and hyper-parameters that enhance PointGoal navigation performance under constrained resources.

Findings

Performance improved by up to 38% on Gibson and 220% on Matterport3D.

Key design choices include advantage estimation, visual encoder architecture, and hyper-parameters.

Extensive experiments totaling over 50,000 GPU-hours support the findings.

Abstract

PointGoal navigation has seen significant recent interest and progress, spurred on by the Habitat platform and associated challenge. In this paper, we study PointGoal navigation under both a sample budget (75 million frames) and a compute budget (1 GPU for 1 day). We conduct an extensive set of experiments, cumulatively totaling over 50,000 GPU-hours, that let us identify and discuss a number of ostensibly minor but significant design choices -- the advantage estimation procedure (a key component in training), visual encoder architecture, and a seemingly minor hyper-parameter change. Overall, these design choices to lead considerable and consistent improvements over the baselines present in Savva et al. Under a sample budget, performance for RGB-D agents improves 8 SPL on Gibson (14% relative improvement) and 20 SPL on Matterport3D (38% relative improvement). Under a compute budget, performance for RGB-D agents improves by 19 SPL on Gibson (32% relative improvement) and 35 SPL on Matterport3D (220% relative improvement). We hope our findings and recommendations will make serve to make the community's experiments more efficient.