Parasara Sridhar Duggirala

Manish Goyal, Miheer Dewaskar, Parasara Sridhar Duggirala

We propose a new technique for performing state space exploration of closed loop control systems with neural network feedback controllers. Our approach involves approximating the sensitivity of the trajectories of the closed loop dynamics. Using such an approximator and the system simulator, we present a guided state space exploration method that can generate trajectories visiting the neighborhood of a target state at a specified time. We present a theoretical framework which establishes that our method will produce a sequence of trajectories that will reach a suitable neighborhood of the target state. We provide thorough evaluation of our approach on various systems with neural network feedback controllers of different configurations. We outperform earlier state space exploration techniques and achieve significant improvement in both the quality (explainability) and performance (convergence rate). Finally, we adopt our algorithm for the falsification of a class of temporal logic specification, assess its performance against a state-of-the-art falsification tool, and show its potential in supplementing existing falsification algorithms.

Arkaprava Gupta, Nicholas Carter, William Zellers et al.

Deep Brain Stimulation (DBS) has proven to be a promising treatment of Parkinson's Disease (PD). DBS involves stimulating specific regions of the brain's Basal Ganglia (BG) using electric impulses to alleviate symptoms of PD such as tremors, rigidity, and bradykinesia. Although most clinical DBS approaches today use a fixed frequency and amplitude, they suffer from side effects (such as slurring of speech) and shortened battery life of the implant. Reinforcement learning (RL) approaches have been used in recent research to perform DBS in a more adaptive manner to improve overall patient outcome. These RL algorithms are, however, too complex to be trained in vivo due to their long convergence time and requirement of high computational resources. We propose a new Time & Threshold-Triggered Multi-Armed Bandit (T3P MAB) RL approach for DBS that is more effective than existing algorithms. Further, our T3P agent is lightweight enough to be deployed in the implant, unlike current deep-RL strategies, and even forgoes the need for an offline training phase. Additionally, most existing RL approaches have focused on modulating only frequency or amplitude, and the possibility of tuning them together remains greatly unexplored in the literature. Our RL agent can tune both frequency and amplitude of DBS signals to the brain with better sample efficiency and requires minimal time to converge. We implement an MAB agent for DBS for the first time on hardware to report energy measurements and prove its suitability for resource-constrained platforms. Our T3P MAB algorithm is deployed on a variety of microcontroller unit (MCU) setups to show its efficiency in terms of power consumption as opposed to other existing RL approaches used in recent work.

Bineet Ghosh, Sandeep Chinchali, Parasara Sridhar Duggirala

A robot can invoke heterogeneous computation resources such as CPUs, cloud GPU servers, or even human computation for achieving a high-level goal. The problem of invoking an appropriate computation model so that it will successfully complete a task while keeping its compute and energy costs within a budget is called a model selection problem. In this paper, we present an optimal solution to the model selection problem with two compute models, the first being fast but less accurate, and the second being slow but more accurate. The main insight behind our solution is that a robot should invoke the slower compute model only when the benefits from the gain in accuracy outweigh the computational costs. We show that such cost-benefit analysis can be performed by leveraging the statistical correlation between the accuracy of fast and slow compute models. We demonstrate the broad applicability of our approach to diverse problems such as perception using neural networks and safe navigation of a simulated Mars rover.

Abolfazl Karimi, Manish Goyal, Parasara Sridhar Duggirala

In this paper, we perform safety and performance analysis of an autonomous vehicle that implements reactive planner and controller for navigating a race lap. Unlike traditional planning algorithms that have access to a map of the environment, reactive planner generates the plan purely based on the current input from sensors. Our reactive planner selects a waypoint on the local Voronoi diagram and we use a pure-pursuit controller to navigate towards the waypoint. Our safety and performance analysis has two parts. The first part demonstrates that the reactive planner computes a plan that is locally consistent with the Voronoi plan computed with full map. The second part involves modeling of the evolution of vehicle navigating along the Voronoi diagram as a hybrid automata. For proving the safety and performance specification, we compute the reachable set of this hybrid automata and employ some enhancements that make this computation easier. We demonstrate that an autonomous vehicle implementing our reactive planner and controller is safe and successfully completes a lap for five different circuits. In addition, we have implemented our planner and controller in a simulation environment as well as a scaled down autonomous vehicle and demonstrate that our planner works well for a wide variety of circuits.