This paper presents a novel terrain-adaptive local trajectory planner designed for the autonomous operation of off-road vehicles on deformable terrains. State-of-the-art solutions either do not account for deformable terrains, or do not offer sufficient robustness or computational speed. To bridge this research gap, the paper introduces a novel model predictive control (MPC) formulation. In contrast to the prevailing state-of-the-art approaches that rely exclusively on hard or soft constraints for obstacle avoidance, the present formulation enhances robustness by incorporating both types of constraints. The effectiveness and robustness of the formulation are evaluated through extensive simulations, encompassing a wide range of randomized scenarios, and compared against state-of-the-art methods. Subsequently, the formulation is augmented with an optimal-control-oriented terramechanics model from the literature, explicitly addressing terrain deformation. Additionally, a terrain estimator employing the unscented Kalman filter is utilized to dynamically adjust the sinkage exponent online, resulting in a terrain-adaptive formulation. This formulation is tested on a physical vehicle in real world experiments against a rigid-terrain formulation as the benchmark. The results showcase the superior safety and performance achieved by the proposed formulation, underscoring the critical significance of integrating terramechanics knowledge into the planning process. Specifically, the proposed terrain-adaptive formulation achieves reduced mean absolute sideslip angle, decreased mean absolute yaw rate, shorter time to goal, and a higher success rate, primarily attributed to its enhanced understanding of terramechanics within the planner.

Workload management is a critical concern in shared control of unmanned ground vehicles. In response to this challenge, prior studies have developed methods to estimate human operators’ workload by analyzing their physiological data. However, these studies have primarily adopted a single-model-single-feature or a single-model-multiple-feature approach. The present study proposes a Bayesian inference model to estimate workload, which leverages different machine learning models for different features. We conducted a human subject experiment with 24 participants, in which a human operator teleoperated a simulated High Mobility Multipurpose Wheeled Vehicle (HMMWV) with the help from an autonomy while performing a surveillance task simultaneously. Participants’ eye-related features, including gaze trajectory and pupil size change, were used as the physiological input to the proposed Bayesian inference model. Results show that the Bayesian inference model achieves a 0.823 F1 score, 0.824 precision, and 0.821 recall, outperforming the single models.

In this article, we develop a modular neural network for real-time (>10 Hz) semantic mapping in uncertain environments, which explicitly updates per-voxel probabilistic distributions within a neural network layer. Our approach combines the reliability of classical probabilistic algorithms with the performance and efficiency of modern neural networks. Although robotic perception is often divided between modern differentiable methods and classical explicit methods, a union of both is necessary for real-time and trustworthy performance. We introduce a novel convolutional Bayesian kernel inference (ConvBKI) layer which incorporates semantic segmentation predictions online into a 3-D map through a depthwise convolution layer by leveraging conjugate priors. We compare ConvBKI against state-of-the-art deep learning approaches and probabilistic algorithms for mapping to evaluate reliability and performance. We also create a robot operating system package of ConvBKI and test it on real-world perceptually challenging off-road driving data.

Robotic perception is currently at a cross-roads between modern methods, which operate in an efficient latent space, and classical methods, which are mathematically founded and provide interpretable, trustworthy results. In this paper, we introduce a Convolutional Bayesian Kernel Inference (Con-vBKI) layer which learns to perform explicit Bayesian inference within a depthwise separable convolution layer to maximize efficency while maintaining reliability simultaneously. We apply our layer to the task of real-time 3D semantic mapping, where we learn semantic-geometric probability distributions for LiDAR sensor information and incorporate semantic predictions into a global map. We evaluate our network against state-of-the-art semantic mapping algorithms on the KITTI data set, demonstrating improved latency with comparable semantic label inference results.

This article presents a reliable and robust rapidly exploring random tree (R2-RRT*) algorithm to tackle challenges in mission planning of off-road autonomous ground vehicles (AGVs) under uncertain terrain environment. Two types of mobility reliability metrics, namely state mobility reliability (SMR) and mission mobility reliability (MMR), are first defined to quantify the mobility reliability of an AGV and to incorporate mobility reliability into mission planning. SMR measures the probability that a vehicle can pass through a specific location on a map of interest, whereas MMR quantifies mobility reliability of a mission path with the consideration of dependence of soil properties and slope over space. Based on the defined SMR and MMR metrics, two reliability-based robust mission planning models are developed to identify optimal paths that have robust travel time and satisfy specific reliability requirements. Moreover, a reliability-based path smoothing algorithm is developed to address the suboptimality of R2-RRT*. Results of a case study demonstrate the efficacy of the proposed models and algorithms.Note to Practitioners—This article was motivated to explicitly account for the uncertain terrain environment in mission planning of off-road autonomous ground vehicles (AGVs). Existing approaches, e.g., RRT and its variants, in general, oversimplify the uncertainty sources and overlook the reliability of vehicle mobility. These simplifications could lead to failure (i.e., immobility) of off-road AGVs on the obtained paths. This article suggests a reliability-based mission planning model by incorporating the proposed SMR and MMR metrics into mission planning considering the spatial-dependent uncertainty sources. To identify a reliable and robust path, we extend RRT* to R2-RRT* that achieves a tradeoff between mission cost and reliability. The proposed reliability-based mission planning model, however, is not limited to RRT* and can also be integrated with other path planning algorithms. It can be also applied to other unmanned vehicles and robots, such as the motion planning of unmanned aerial vehicles (UAVs) in adverse weather conditions.

Mobility of teleoperated unmanned ground vehicles can be significantly compromised under large communication delays, if the delays are not compensated. This article considers a recently developed delay compensation theory and presents its first empirical evaluation in improving mobility and drivability of a high-speed teleoperated vehicle under large delays. The said delay compensation theory is a predictor-based framework. Two realizations of this framework are considered: a model-free realization that relies only on model-free predictors, and a blended realization, where the heading predictions from the model-free predictor are blended with those from a steering-model-based feedforward predictor for a more accurate prediction of the vehicle heading. A teleoperated track-following task is designed in a human-in-the-loop simulation platform. This platform is used to compare the teleoperation performance with and without the predictor-based framework under both constant and varying delays. Through repeated measurement analysis of variance, it is concluded that the predictor-based framework is effective in achieving a higher vehicle speed, more accurate lateral control, and better drivability as indicated by the three performance metrics of track completion time, track keeping error, and steering control effort, respectively. In addition, it is shown that the blended architecture can lead to further improvements in these metrics compared to using the model-free predictors alone. The analysis also shows that there is no statistically significant difference between constant and varying delay cases in the designed experiment, nor there is any direct relation between drivers' skill level and level of improvement in metrics.