MetaCropFollow: Few-Shot Adaptation with Meta-Learning for Under-Canopy Navigation

TL;DR

This paper trains a base-learner that can quickly adapt to new conditions, enabling more robust navigation in low-data regimes, and explores the use of Meta-Learning to overcome this domain shift using a minimal amount of data.

Abstract

Autonomous under-canopy navigation faces additional challenges compared to over-canopy settings - for example the tight spacing between the crop rows, degraded GPS accuracy and excessive clutter. Keypoint-based visual navigation has been shown to perform well in these conditions, however the differences between agricultural environments in terms of lighting, season, soil and crop type mean that a domain shift will likely be encountered at some point of the robot deployment. In this paper, we explore the use of Meta-Learning to overcome this domain shift using a minimal amount of data. We train a base-learner that can quickly adapt to new conditions, enabling more robust navigation in low-data regimes.

Figures (4)

• Figure 1: Visual under-canopy agricultural navigation is challenging due to large variations in lighting, appearance of the crops and soil throughout the season. We evaluate the use of MAML for few-shot adaptation of a semantic keypoint prediction network introduced by Sivakumar-RSS-24.
• Figure 2: Samples of the dataset used.
• Figure 3: Our process starts by taking a subset of the data to train three models: Non-MAML, MAML, ANIL. For each of those, the validation set is from the same domain as the training set. We then see how they perform in the three domains (early, late and very late) calculating a test loss for each.
• Figure 4: The test results of a model trained only on early-season data with MAML and without. Both models work good for early season, but only the MAML-trained model is capable of adapting to other seasons.