Anusha Srikanthan

I'm a 4th year PhD student at the GRASP Lab, University of Pennsylvania in Philly, where I'm jointly advised by Nikolai Matni, an expert mathematician and Vijay Kumar, an expert roboticist.

I'm motivated by problems that require rigorous mathematical guarantees while being practically useful for robot applications. My previous research was on the topic of guaranteeing dynamic feasibility in trajectory optimization for nonlinear systems via optimization theory, with applications to under-actuated robots. I've received the Excellent Paper Award for my Workshop paper on Resilient Task Allocation for Heterogeneous Robots at IROS 2021.

Email  /  CV  /  Bio  /  Scholar  /  LinkedIn  /  Github

Research

I'm interested in the problem of safety in robotics using layered control, optimization theory, and machine learning. Most of my research is about discovering mathematical principles governing the design of decision-making hierarchies in task planning, trajectory design, and feedback control. Some papers are highlighted below.

To tackle the combined challenges of state and input constaint satisfaction, dynamic feasibility, safety, and real-time computation, we present a layered control architecture (LCA) consisting of an offline path library generation layer, and an online path selection and safety layer.

This paper proposes a suboptimal MPC scheme based on the alternating direction method of multipliers (ADMM). We show that using a warm-start approach combined with enough iterations per time-step, yields an ADMM-based suboptimal MPC scheme which asymptotically stabilizes the system and maintains recursive feasibility.

We propose the use of alternating direction method of multipliers algorithm (ADMM) on nonlinear optimal control problems to derive a layered control architecture.

Motivated by the lack of existing methods to account for controller cost in trajectory planning for robotic systems, we propose a principle derivation to decompose a nonlinear optimal control problem into trajectory generation and feedback control layers.

Task allocation in multi-robot teams is often hindered by unknown task reward functions. This work introduces the COCOA problem, addressed by a continuous-armed bandit algorithm, which uses online optimization to form coalitions that maximize unknown task rewards while respecting resource constraints in real time.

This work proposes a two-part framework that infers heterogeneous strategies from expert demonstrations and adaptively selects the best strategy for coalition formation based on a team's capabilities. Through numerical simulations, StarCraft II battles, and multi-robot emergency-response tasks, the framework outperforms existing approaches in requirement satisfaction, resource utilization, and task success rates.

This thesis contributes two frameworks to learn implicit task requirements directly from expert demonstrations of coalition formation.

This paper proposes a resilient task allocation framework for heterogeneous robot teams operating in dynamic environments. Our approach enables robots to adapt to failures and disturbances, improving task success rates. This work was awarded the Excellent Paper Award at the IROS 2021 Workshop.

Industry Experience

• Identified the challenges of lack of interpretability of model outputs and label quality in their current framework.
• Proposed a new loss function based on task-specific metrics from observing the correlation trends of metrics through experiments.
• Conducted performance evaluations of the trained models to improve the interpretability of the scoring pipeline. Link to report.

• Designed and implemented a Safety Duplication Plugin for multiple error detection using concepts of redundancy and clock domains and integrated it on Perforce using Perl scripts with Viva embedded code programmed on a UNIX based OS.
• Formalized hierarchical changes in the internal architecture of the IP module for making it plugin compatible, increasing the safety compliance at the hardware level to prevent failure when the chip is used in self-driving cars. Link to project report.

Software projects

Invited Talks and Teaching

Education