Saurav Agarwal
Postdoctoral Researcher
GRASP Laboratory, University of Pennsylvania
Contact: SauravAg@seas.upenn.edu

Research Focus: Decentralized and Collaborative Intelligent Robotic Systems

Research Interests

My research explores the development of large-scale decentralized and collaborative intelligent systems, enabling robot teams to operate cohesively through advanced environmental perception, inter-robot communication, and coordinated actions. By combining learning-based approaches, robust theoretical frameworks, and practical field implementations, I aim to enhance scalability, efficiency, resilience, and performance for complex robotic tasks. Key technologies in my work include graph neural networks (GNNs), constrained learning, and optimization frameworks.

My Ph.D. research introduced a unified framework for generalized coverage of point, curve, and area features in environments, formalized as graph-based optimization problems. This work led to the design of approximation algorithms with provable guarantees and heuristics for fast, large-scale applications, validated through extensive simulations and real-world experiments. Before my Ph.D., I focused on analyzing and designing mechanisms and parallel robots using optimization techniques. My expertise also spans symbolic algebra systems, open-source library development, project leadership, and student mentorship.

LPAC: Learnable Perception-Action-Communication Loops with Applications to Coverage Control

A canonical learnable perception-action-communication (LPAC) architecture for decentralized and collaborative intelligent systems. (1) Perception: a CNN processes localized observations and generates an abstract representation.  (2) Communication: a graph neural network (GNN) computes what information to share and how to integrate the information received from other robots. (3) Action: a shallow MLP computes control actions based on the GNN output. The three modules are executed on each robot independently, with the GNN facilitating collaboration.

The LPAC model scales well to collaborative systems with larger numbers of robots and features without further retraining or fine-tuning.

Coverage Control Problem
Collaboratively provide sensor coverage to a convex environment. The robots have limited sensing and communication capabilities.

Robust Generalization and Transfer
Demonstrates strong performance across diverse environments, larger swarms, and noisy robot position estimates. The learned policy scales without retraining or performance degradation.

Real-World Applicability
Validates on a real-world traffic light dataset, showing that the trained LPAC model transfers well to practical scenarios, indicating promise for large-scale, decentralized swarm deployments.

Related Publications:

Coverage of Linear Features using Multiple Robots



Related Publications:

Variable Formation

Development of algorithms to simultaneously compute the optimal assignments and formation parameters for a team of robots from a given initial formation to a variable goal formation.

Publication:
S. Agarwal and S. Akella, "Simultaneous optimization of assignments and goal formations for multiple robots," IEEE International Conference on Robotics and Automation, Brisbane, Australia, May 2018.
PDF       bib     Video