Profile photo

Saurav Agarwal

Postdoctoral Researcher
GRASP Laboratory, University of Pennsylvania
Contact: SauravAg [at] upenn [dot] edu
Research Focus:
Decentralized and Collaborative Intelligent Robotic Systems

My research interests include 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

What to communicate with neighboring agents
How to integrate received information with own localized observations
Decentralized communication and computation
Same policy executed independently on each agent

Architecture for Learnable Perception-Action-Communication Loops
A canonical lpac architecture for decentralized and collaborative intelligent systems. (1) Perception: a cnn processes localized observations and generates an abstract representation. (2) Communication: a gnn computes what information to share and how to integrate the information received from other robots. (3) Action: a shallow mlp computes control actions on the gnn output. The three modules are executed on each robot independently, with the gnn facilitating collaboration.

Heatmap showing generalization of LPAC

Coverage Control Problem:
Collaboratively provide sensor coverage with robots that have limited sensing and communication capabilities. Every robot runs the same policy independently based on their own localized observations and nearby communications.

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

  • PDF|IEEE|arXiv
    S. Agarwal, R. Muthukrishnan, W. Gosrich, V. Kumar, and A. Ribeiro IEEE Transactions on Robotics (T-RO), vol. 41, October 2025
    Coverage control is the problem of navigating a robot swarm to collaboratively monitor features or a phenomenon of interest not known a priori. The problem is challenging in decentralized settings with robots that have limited communication and sensing capabilities. We propose a learnable Perception-Action-Communication (LPAC) architecture for the problem, wherein a convolutional neural network (CNN) processes localized perception; a graph neural network (GNN) facilitates robot communications; finally, a shallow multi-layer perceptron (MLP) computes robot actions. The GNN enables collaboration in the robot swarm by computing what information to communicate with nearby robots and how to incorporate received information. Evaluations show that the LPAC models—trained using imitation learning—outperform standard decentralized and centralized coverage control algorithms. The learned policy generalizes to environments different from the training dataset, transfers to larger environments with more robots, and is robust to noisy position estimates. The results indicate the suitability of LPAC architectures for decentralized navigation in robot swarms to achieve collaborative behavior.
  • PDF|IEEE|arXiv
    J. Cerviño, S. Agarwal, V. Kumar, and A. Ribeiro IEEE International Conference on Robotics and Automation (ICRA), May 2025
    The multi-objective coverage control problem requires a robot swarm to collaboratively provide sensor coverage to multiple heterogeneous importance density fields (IDFs) simultaneously. We pose this as an optimization problem with constraints and study two different formulations: (1) Fair coverage, where we minimize the maximum coverage cost for any field, promoting equitable resource distribution among all fields; and (2) Constrained coverage, where each field must be covered below a certain cost threshold, ensuring that critical areas receive adequate coverage according to predefined importance levels. We study the decentralized setting where robots have limited communication and local sensing capabilities, making the system more realistic, scalable, and robust. Given the complexity, we propose a novel decentralized constrained learning approach that combines primal-dual optimization with a Learnable Perception-Action-Communication (LPAC) neural network architecture. We show that the Lagrangian of the dual problem can be reformulated as a linear combination of the IDFs, enabling the LPAC policy to serve as a primal solver. We empirically demonstrate that the proposed method (i) significantly outperforms state-of-the-art decentralized controllers by 30% on average in terms of coverage cost, (ii) transfers well to larger environments with more robots, and (iii) is scalable in the number of IDFs and robots in the swarm.

Coverage of Linear Features using Multiple Robots

How to survey a linear infrastructures like road networks and powerlines?
Path planning for robot teams with limited battery life
Operating modes: Servicing (taking images) and Deadheading (fast travel)
Asymmetric costs and demands results in APX-hard problem

Line coverage visualization

  • PDF|IEEE|arXiv
    S. Agarwal and S. Akella IEEE Transactions on Robotics (T-RO), vol. 40, January 2024
    The line coverage problem involves finding efficient routes for the coverage of linear features by one or more resource-constrained robots. Linear features model environments such as road networks, power lines, and oil and gas pipelines. Two modes of travel are defined for robots: servicing and deadheading. A robot services a feature if it performs task-specific actions, such as taking images, as it traverses the feature; otherwise, it is deadheading. Traversing the environment incurs costs (e.g., travel time) and resource demands (e.g., battery life). Servicing and deadheading can have different cost and demand functions, which can be direction dependent. The environment is modeled as a graph, and an integer linear program is provided. As the problem is NP-hard, we design a fast and efficient heuristic algorithm, Merge-Embed-Merge (MEM). Exploiting the constructive property of the MEM algorithm, algorithms for line coverage of large graphs with multiple depots are developed. Furthermore, turning costs and nonholonomic constraints are efficiently incorporated into the algorithm. The algorithms are benchmarked on 100 road networks and demonstrated in experiments with aerial robots.
  • PDF|Networks|arXiv
    S. Agarwal and S. Akella Networks, vol. 82, July 2023
    Line coverage is the task of servicing a given set of one-dimensional features in an environment. It is important for the inspection of linear infrastructure such as road networks, power lines, and oil and gas pipelines. This paper addresses the single robot line coverage problem for aerial and ground robots by modeling it as an optimization problem on a graph. The problem belongs to the broad class of arc routing problems and is closely related to the rural postman problem (RPP) on asymmetric graphs. The paper presents an integer linear programming formulation with proofs of correctness. Using the minimum cost flow problem, we develop approximation algorithms with guarantees on the solution quality. These guarantees also improve the existing results for the asymmetric RPP. The main algorithm partitions the problem into three cases based on the structure of the required graph, i.e., the graph induced by the features that require servicing. We evaluate our algorithms on road networks from the 50 most populous cities in the world, consisting of up to 730 road segments. The algorithms, augmented with improvement heuristics, run within 3s and generate solutions that are within 10% of the optimum. We experimentally demonstrate our algorithms with commercial UAVs.

Formations: Optimal Assignment and Transformation

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.

The shape of goal formation is provided as input
The scale and location parameters for the goal formation are optimized
Optimal assignments of the robots to the goal positions
The sum of squared travel distance is minimized
Guaranteed collision-free trajectories
Robots start simultaneously and reach their goal positions simultaneously
\(O(n^3)\) running time complexity
Checkout DoodleBots for a live demo


PDF|IEEE|DoodleBots
S. Agarwal and S. Akella IEEE International Conference on Robotics and Automation (ICRA), May 2018
This paper presents 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 (where the shape of the goal formation is given, and its scale and location parameters must be optimized). We assume the \( n \) robots are identical spheres. We use the sum of squared travel distances as the objective function to be minimized, which also ensures that the trajectories are collision free. We show that this assignment with variable goal formation problem can be transformed to a linear sum assignment problem (LSAP) with pseudo costs that we establish are independent of the formation parameters. The transformed problem can then be solved using the Hungarian algorithm in \( \mathcal O(n^3) \) time. Thus the assignment problem with variable goal formations using this new approach has the same \( \mathcal O(n^3) \) time complexity as the standard assignment problem with fixed goal formations. Results from simulations on 200 and 600 robots are presented to show the algorithm is sufficiently fast for practical applications.