We present a self-supervised approach for learning to predict traversable paths for wheeled mobile robots that require good traction to navigate. Our algorithm, termed WayFAST (Waypoint Free Autonomous Systems for Traversability), uses RGB and depth data, along with navigation experience, to autonomously generate traversable paths in outdoor unstructured environments. Our key inspiration is that traction can be estimated for rolling robots using kinodynamic models. Using traction estimates provided by an online receding horizon estimator, we are able to train a traversability prediction neural network in a self-supervised manner, without requiring heuristics utilized by previous methods. We demonstrate the effectiveness of WayFAST through extensive field testing in varying environments, ranging from sandy dry beaches to forest canopies and snow covered grass fields. Our results clearly demonstrate that WayFAST can learn to avoid geometric obstacles as well as untraversable terrain, such as snow, which would be difficult to avoid with sensors that provide only geometric data, such as LiDAR. Furthermore, we show that our training pipeline based on online traction estimates is more data-efficient than other heuristic-based methods.

This article presents a deep learning-based model predictive control (MPC) algorithm for control affine nonlinear discrete-time systems with matched and bounded state-dependent uncertainties of unknown structure. Since the structure of uncertainties is not known, a deep neural network is employed to approximate them. In order to avoid any unwanted behavior during the learning phase, a tube-based nonlinear model predictive controller is employed, which ensures satisfaction of constraints and input-to-state stability of the closed-loop states. In addition, the proposed approach guarantees the convergence of states to the origin under certain conditions. To make the algorithm implementable online, a dual-timescale adaptation mechanism is utilized, where the weights of the output layer of the neural network are updated each time instant using a weight update law, while the inner layers are repeatedly trained in self-supervised manner by using the adaptive actions as labels for the training. Our results are validated through a numerical experiment, which indicates that the proposed deep MPC architecture is effective in learning to control safety critical systems without suffering instability drawbacks.

Compact robots operating beneath the crop canopy present great potential for a range of autonomous and remote tasks, including phenotyping, soil analysis, and cover cropping. Under-canopy navigation presents unique challenges, such as the need for a navigation system that can traverse diverse crop types, navigate despite sensory obstructions, and manage sensory noise effectively. Aiming to solve this problem in a scalable manner, we present a novel navigation method that uses a self-supervised neural network tailored for row-following in under-canopy plantations for mobile robots. Our method, termed CROW (Crop-ROW navigation), integrates perception, waypoint generator, and control components, and is capable of handling variations in luminosity, topology, types of plantations, and plant growth stages. By using a Deep Learning-based approach to interpret LiDAR scans, we convert the detected rows of crops into lines, establishing waypoints for the controller based on fundamental geometric principles. To address the computational complexity inherent in standard Model Predictive Controller solvers, we employ a Constrained Iterative Linear Quadratic approach. Our system has been validated in both simulated and real-world environments, demonstrating successful navigation through 115-meter corn rows with little to no intervention, i.e., requiring only \(3 \pm 3\) interventions per row experiment.

Small robotic vehicles have been navigating agricultural fields in the pursuit of new possibilities to increase agricultural production and to meet the increasing food and energetic demands. However, a perception system with reliable awareness of the surroundings remains a challenge to achieve autonomous navigation. Camera and single-layer laser scanners have been the primary sources of information, yet the first suffers from outdoor light sensibility and both from occlusion by leaves. This paper describes a three-dimensional system acquisition for corn crops. The sensing core is a single-layer UTM30-LX laser scanner rotating around its axis, while an inertial sensor provides angular measurements. With the rotation, multiple layers are used to compose a 3D point cloud, which is represented by a two-dimensional occupancy grid. Each cell is filled according to the number of readings, and their weights derive from two procedures: firstly, a mask enhances vertical entities (stalks); secondly, two Gaussian functions on the expected position of the immediate neighboring rows weaken readings in the middle of the lane and farther rows. The resulting occupancy grid allows the representation of the cornrows by virtual walls, which are used as references to a wall follower algorithm. According to experimental results, the virtual walls are segmented with reduced influence from straying leaves and sparse weeds when compared to the segmentation done with single-layer laser scanner data. Indeed, 64.02% of 3D outputs are within 0.05 m limit error from expected lane width, while only 11.63% of single-layer laser data are within same limit.