Myoelectric control systems for assistive devices are still unreliable. The user's input signals can become unstable over time due to e.g. fatigue, electrode displacement, or sweat. Hence, such controllers need to be constantly updated and heavily rely on user feedback. In this paper, we present an automatic failure detection method which learns when plausible predictions become unreliable and model updates are necessary.Our key insight is to enhance the control system with a set of generative models that learn sensible behaviour for a desired task from human demonstration. We illustrate our approach on a grasping scenario in Virtual Reality, in which the user is asked to grasp a bottle on a table. From demonstration our model learns the reach-to-grasp motion from a resting position to two grasps (power grasp and tridigital grasp) and how to predict the most adequate grasp from local context, e.g. tridigital grasp on the bottle cap or around the bottleneck. By measuring the error between new grasp attempts and the model prediction, the system can effectively detect which input commands do not reflect the users intention. We evaluated our model in two cases: i) with both position and rotation information of the wrist pose, and ii) with only rotational information. Our results show that our approach detects statistically highly significant differences in error distributions with p < 0.001 between successful and failed grasp attempts in both cases.

Plant phenotyping is an essential step in the plant breeding cycle, necessary to ensure food safety for a growing world population. Standard procedures for evaluating three-dimensional plant morphology and extracting relevant phenotypic characteristics are slow, costly, and in need of automation. Previous work towards automatic semantic segmentation of plants relies on explicit prior knowledge about the species and sensor set-up, as well as manually tuned parameters. In this work, we propose to use a supervised machine learning algorithm to predict per-point semantic annotations directly from point cloud data of whole plants and minimise the necessary user input. We train a PointNet++ variant on a fully annotated procedurally generated data set of partial point clouds of tomato plants, and show that the network is capable of distinguishing between the semantic classes of leaves, stems, and soil based on structural data only. We present both quantitative and qualitative evaluation results, and establish a proof of concept, indicating that deep learning is a promising approach towards replacing the current complex, laborious, species-specific, state-of-the-art plant segmentation procedures.

Essential to agricultural robot deployment in farms are accurate topological maps, which are manually created in current systems. In this work we present a novel approach to automatically generate a topological map along crop rows from aerials images for the deployment of agricultural mobile robots. We evaluate our system in a digital twin of a farm environment using real-world textures and physical simulation, and also demonstrate its applicability to aerial images of a real farm.

No abstract