2022 ARC Research Seminar - Fall Series

PI: Dr. Yue Wang (Clemson U.)
project link

In this talk, we present a trust-based active learning framework to infer human trust in multi-robot systems (MRS) and encode it into a human-multi-robot collaborative bounding overwatch task in offroad environments. We first develop (1) a linear state space model to quantify the causality between human trust in MRS and the robots' perceived environment attributes, such as traversability and visibility; and (2) a linear temporal logic (LTL) formula to impose the correct behaviors of MRS. We then shape a reward function for the MRS to plan trajectories by referring to the linear state space trust model and LTL formula. Furthermore, we rely on Bayesian inference and Markov Chain Monte Carlo to learn the parameters of the reward function. The parameter learning process goes concurrently with the trajectory planning process under a trust-based exploration strategy. Our simulation in ROS Gazebo shows that the optimal trajectories along the episodes of active learning can finally stabilize on one or several trajectories. It proves the usability of our trust-based active learning in inferring human trust dynamics and finding the optimal trajectory in the human-multi-robot bounding overwatch.

PI: Dr. Manish Bansal (Virginia Tech)
project link

Interdiction problems are characterized as games played between two players: an interdictor/attacker/leader and defender/follower. The leader is a player who makes interdiction decisions using limited resources to degrade the follower’s performance, and the follower makes decisions after observing the interdiction decision. We introduce new algebraic modeling frameworks that allow uncertainty in the success and impact of the attacks, adjustments based on risk-appetite (risk-receptive or risk-averse) of the decision makers, and incomplete information of probability distribution associated with uncertain data. We present reformulation and cutting-plane based exact methods for solving these models. Specifically, we consider two games: (a) Shortest Path Network Interdiction, where the follower finds a minimum cost path between a given pair of source and destination nodes of the interdicted network, and (b) Adversarial Camera View-Frame Placement, where the follower determines pan/tilt/zoom of multiple cameras to determine subregions of maximum priority based on input provided by a set of agents located in an adversarial environment. This is joint work with PhD students: Sumin Kang and Sunghoon Park.

PI: Dr. Bogdan Epureanu (U. of Michigan)
project link

Coordination in multi-agent systems is a vital component in teaming effectiveness, from swarm robots to human-machine interactions. While multi-agent reinforcement learning shapes team behaviors from experience, interpreting learned coordination strategies offer benefits in understanding complex agent dynamics and further improvement in developing adaptive strategies for evolving and unexpected situations. Depending on the complexity of the environment, agent decisions might depict emergent coordination strategies from following predetermined rules to exploiting incentive-driven policies. In this work, we develop an approach to quantitatively measure team coordination by collecting decision time series data, detecting causality between agents, and identifying statistically high coordinated sub-teams. This autonomously emergent team coordination is then analyzed using rule-based and reinforcement learning-based strategies in different demand situations (stress levels) in a disaster-relief scenario. We demonstrate that choosing the proper coordination pattern at different stages of an operation is key to successful teaming without catastrophic failures, and we propose a metric to quantify performance. The proposed data-driven method provides opportunities to reveal how small communities (high coordinated sub-teams) or leader-follower (one directional causality) relations are autonomously forming and influencing the overall team performance.

PI: Dr. Shravan Veerapaneni (U. of Michigan)
project link

In this talk, we will review some of the recent advances by our team in developing and applying variational quantum algorithms to various hard computational problems arising in autonomous ground vehicle systems research. Specifically, we will discuss quantum acceleration of second-order solvers for the differential variational inequality problem and applying quantum approximate optimization algorithm (QAOA) to the clustering problem arising in autonomous mobility tasks. A warm-start strategy for accelerating the QAOA will be presented along with numerical results showcasing runs on IBM quantum hardware.

PI: Dr. Andrea L'Afflitto, Dr. Corina Sandu (Virginia Tech)
project link

Assets for the US Army are autonomous ground vehicles (UGVs), i.e., ground vehicles able to operate at a high-speed off-road and without human supervision. To be useful in contested scenarios, these vehicles must be deployable in GPS-denied or GPS-contested environments and be able to cope with adversaries, whose locations and capabilities are unknown a priori. This presentation will present some of the key features of a project funded by ARC since 2021, which is aimed at enabling off-road maneuvers at high-speed using a terrain-aware guidance system. Unique features of the proposed guidance system are the user’s ability to set the mission optempo (stealthy vs. forceful and rapid vs. deliberate) before the mission starts. Thus, according to this input, the proposed guidance system decides to what extent obstacles can be employed to shelter the vehicle and how speed should be regulated as a safety mechanism while meeting the mission objectives. An additional key feature of the proposed system is the live feedback from an ad-hoc terrain estimation system, whose data are pivotal for the guidance system to make decisions in consideration of the terrain properties.

PI: Dr. Zissimos P. Mourelatos (Oakland U.), Dr. Zhen Hu (U. of Michigan - Dearborn)
project link

Off-road autonomous ground vehicles (AGVs) may fail to complete a mission in unstructured off-road environments for many reasons, such as running out of power, overturning, loss of mobility, etc. While multi-physics simulations play a vital role in predicting some failure modes, they are computationally expensive to be used in path planning, especially in the presence of uncertainty. Instead, data-driven mobility models are usually employed to replace the high-fidelity simulation models. Even though data-driven mobility models can drastically reduce the computational effort of mobility prediction, they may lead to wrong decisions if the models are not properly trained. In this talk, we will present our progress in developing data-driven mobility models using synthetic data generated from high-fidelity simulations. To demonstrate our developments, we use the vehicle maximum attainable speed and vertical acceleration, as an example. Other measures can be also used. Using the developed data-driven mobility model, we will discuss reliable path planning of off-road AGVs subject to multiple failure modes. Finally, we will discuss how uncertainty quantification of data-driven mobility models can impact data-driven mobility model prediction for path planning.

PI: Dr. Wing-Yue Geoffrey Louie (Oakland U.)
project link

Video game environments enable the US army to substitute physical experimentation for virtual experimentation to rapidly obtain feedback from end-users towards technologies and investigate the effects of these technologies on a mission. Virtual experimentation is not just a replacement for studying human-robot interactions in the physical world but can produce insights that may not have been feasible previously. However, current video game environments do not support stakeholders in spectating these environments because they are designed for the needs of the player. There is an open opportunity for improving the spectator experience so that stakeholders can leverage the advantages of virtual experimentation. In this talk, we will discuss tools that we have been developing to improve the ability of stakeholders to spectate experiments and derive insights. We also present preliminary results from a case study conducted to investigate the potential new insights that can be drawn using the developed spectating tools.