2024 ARC Research Seminar - Fall Series

PIs: Dr. Kira Barton, Dr. Chris Vermillion (presenter) (U. of Michigan)
project link

Abstract 1: Driving is characterized by the shared interaction between a vehicle (including its controller), which can be modeled using first-principles techniques, and a human operator whose behavior is best understood through other modeling frameworks (e.g., data-driven models). Making matters more challenging, the optimal blend of human input and input from the automatic control system depends sensitively on vehicle-, terrain-, and mission-specific characteristics. This talk will focus on methods for utilizing cognitive hierarchy theory to characterize the bi-directional interaction between the human operator and underlying vehicle (including its controller), along with iterative learning techniques for deducing the optimal arbitration level between human and autonomous inputs in the context of a driver training simulator environment where a closed course is repeated. The talk will include initial driver-in- the-loop simulation results to illustrate the efficacy of the proposed approaches as well as several of the fundamental research challenges that lie ahead.

PIs: Dr. Shravan Veerapaneni, Dr. James Stokes (U. of Michigan)
project link

Abstract 2: In this talk, we will present an approach for utilizing quantum algorithms to accelerate physics-based terramechanics simulations. The recently-developed photonic quantum chips utilize qumodes as fundamental units of computation (as opposed to qubits), avoid expensive cryogenics in favor of a compact form factor operating at room temperature and are highly scalable, all of which render them as a better fit for tackling the complementarity problem (arising in DEM-C formulations) at scale. By exploiting a Lagrangian formulation involving slack variables, the complementarity problem is mapped to an instance of an constrained quadratic programming problem that can be tackled via the continuous-variable quantum approximate optimization algorithm (CV-QAOA). The implementation of the CV-QAOA relies on the existence of a universal quantum computer, which is specified in terms of a fundamental gate set. We demonstrate how the CV-QAOA can be decomposed into successive application of these gates. Our ongoing efforts to develop a quantum software suite that can be executed on Xanadu’s newest X8 photonic quantum computing chip will be discussed.

PIs: Dr. Daniel Carruth, Dr. Cindy Bethel, Audrey Aldridge (Presenter) (Mississippi State U.)
project link

Abstract for first talk: Information sharing within teams, particularly in teams of human and autonomous agents, significantly impacts team dynamics, including trust, reliance, and overall performance. As autonomous agents are increasingly integrated into human teams, it is critical to understand how sharing of information affects a common understanding of the task (shared mental models) and team function. This study uses a novel configurable virtual testbed to investigate how real-time communication and information availability influence team dynamics. In the testbed, a human and two autonomous agents collaborate on a maze-based search task. During the task, the three team members will try to work together, and the autonomous agents will assess their trust in the human and other agent based on observed task performance. We examine how three levels of available information (no personal or shared information, personal but no shared information, and shared information) affect task performance and the assessment of trust-reliance between the team members (including agent trust of the human participant). Data collection is ongoing, but we expect to find increased information sharing improves team performance and increases trust and reliance on teammates. These findings will have implications for designing communication and information sharing strategies for human-autonomous agent teams. Future work will use the testbed to look at the effects of mismatches in understanding of the tasks (lack of shared mental models) and how our trust-reliance assessments may help identify issues affecting team performance and trust.

PIs: Dr. Dawn Tilbury, Dr. Lionel Robert, Alia Gilbert (Presenter) (U. of Michigan)
project link

Abstract for second talk: While the capacity for autonomy in ground vehicles is continually evolving, enabling these systems to operate independently, they often still require supervision from human operators. Consequently, it is essential to design interfaces that facilitate effective human oversight, allowing operators to assist autonomous systems where both human and machine strengths can be leveraged. This research aims to develop methods to help semi-autonomous UGVs navigating uncertain and complex terrain effectively request and incorporate human input to arrive at their desired destination. When using local path planners, UGVs can get stuck in local minima. We designed an experiment to evaluate intervention methods from a human supervisor if a UGV gets stuck. Participants are asked to watch an interface with first-person views of two UGVs traveling on roadless terrain in a 3D environment and a separate high-level 2D map. UGVs will navigate autonomously through the terrain toward a pre-defined goal location. In the presence of obstacles, the UGVs occasionally get stuck and ask for human intervention. We will evaluate which intervention method is most effective for getting efficiently to the goal while also performing well on a secondary task. The presentation will describe the planned user experiment and expected contributions.

Presented by: Dr. Elena Shrestha; PI: Dr. Ram Vasudevan (U. of Michigan)
project link

Autonomous off-road navigation poses significant challenges due to limited prior knowledge of operational environments (e.g., terrain types, obstacles, and maps), complex vehicle-terrain interactions (e.g., tires on deformable surfaces), sensor noise, and the stochastic nature of real-world conditions. Traditional model-based approaches, like nonlinear model predictive control (NMPC), offer safety guarantees within a limited operational range but fail to generalize to dynamic and uncertain conditions. Data-driven methods, such as reinforcement learning (RL), adapt to changing conditions through interaction data but lack formal safety guarantees and often rely on black-box models. Hybrid policies aim to combine the strengths of both approaches by integrating model-based safety with the adaptability of learning-based methods. However, training learning-based components for off-road scenarios is challenging due to the risks and difficulties of real-world data collection. While simulation-based techniques like domain randomization have shown promise, they often fall short of capturing real-world complexities. By pretraining hybrid off-road driving policies on real-world data, this approach aims to reduce reliance on simulations and improve the safety, adaptability, and robustness of autonomous off-road systems. This ongoing effort aims to enable the use of data collected online to compute trajectories that safely navigate over an unknown terrain through an ensemble of expert policies that can be continuously updated throughout the mission.

Presented by: Mr. Challen Enninful (U. of Michigan)

Achieving real-time receding horizon motion planning for unmanned vehicles with safety guarantees remains a significant challenge, particularly when navigating uncertain terrain with varying payloads. Online trajectory generation methods like Nonlinear Model Predictive Control (NMPC) require solving nonlinear optimizations with fine time discretization to ensure safety, but a fine discretization increases computational demands of the optimization, resulting in a tradeoff between safety and real-time performance. This talk introduces a reachability-based framework to ensure safety amid model uncertainties without sacrificing real-time operation. This method will serve as an expert to train a hybrid RL Mixture of Experts (MoE) policy for the ARC project on high-speed off-road navigation. The framework first bounds uncertainties in vehicle dynamics that stem from different terrain and tire dynamics and varied payloads.Then, zonotope-based reachability analysis is used to compute control-parameterized Forward Reachable Set (FRS) for closed-loop, full-order vehicle dynamics. This pre-computed FRS is integrated into an online optimization framework to ensure safety. Simulation-based evaluation on a full-size vehicle model and hardware tests on a 1/10th race car demonstrate the effectiveness of the proposed method. Compared to state-of-the-art approaches, the proposed method enables vehicles to safely navigate complex environments, showcasing its potential to address real-time safety challenges in motion planning.