Nishan Srishankar

I am currently a Senior AI Research Associate at JPMorgan Chase.

I was previously a Data Scientist/ Machine Learning Engineer 5 at Fidelity AI CoE (Aug21-May23). I graduated with a MSc. in Robotics at WPI working with Prof. Carlo Pinciroli at the Novel Engineering for Swarm Technologies (NEST) Lab. My research involved developing decentralized algorithms for collective spatial perception in a robotic swarm given adversarial conditions and agents. In 2015, I graduated from WPI with a Bachelors in Mechanical Engineering with minors in Aerospace and Electrical Engineering. My undergraduate thesis was to prototype a flapping-wing Micro-Aerial Vehicle.

Previously, I have worked at Honda Research Institute on interpretable relational modeling and at the NASA-SETI Institute on self-supervised knowledge discovery for earth observation data.

My long-term interests are to develop inherently eXplainable multi-robot systems with realistic/online learning curriculums.

Email  |  Resume  |  Github  |  LinkedIn  |  Google Scholar  |  Notes

News
  • [October 2022-] Currently part of an AI/ Advanced AI Alignment reading and project group.
  • [December 2021] Sim2Real paper for documents accepted to NeurIPS21 Data-centric AI workshop.
  • [November 2021] Excited to have gotten a certificate of merit for work done as part of NASA-SETI FDL 2020.
  • [March 2021] Decentralized federated learning (FlowFL) paper for MRS accepted to ICRA 2021.
  • [December 2020] SETI FDL Knowledge Discovery Framework's team's work presented at AGU 2020.
  • [November 2020] Contributed to NASA's Request for Information (Planetary Data Ecosystem) on hurdles faced while building data pipelines and ML models.
  • [October 2020] Explainable Autonomous Driving through Grounded Relational Inference accepted to NeurIPS20 ML4AD (Autonomous Driving) workshop.
  • [August 2020] Showcase of NASA FDL work here.
  • More
Research
  • My thesis research is in developing a decentralized decision-making algorithm in an anonymous swarm of robots that is robust to sensor noise and potential defecting agents whilst being constrained by memory and communication capabilities.

  • I was a Research Programmer on a DARPA-funded Warfighter Analytics for Smartphone Healthcare project to use a smartphone sensor suite to indicate the health of soldiers on the fly.

  • I also worked on a project with the Army Research Laboratory to create a full-stack visual-analytics deep-learning toolbox for real-world corrosion assessment using limited data.
Explainable Autonomous Driving with Grounded Relational Inference
Chen Tang*, Nishan Srishankar*, Sujitha Martin*, Masayoshi Tomizuka*
NeurIPS Machine Learning for Autonomous Driving, 2020
In review, 2021

Abstract | Video

We develop a framework that generates an interpretable object-level interaction graph by grounding the low-dimensional latent space representation on semantic behaviors with expert domain knowledge and apply this to highway driving behavior.

Flow-FL: Data-Driven Federated Learning for Spatio-Temporal Predictions in Multi-Robot Systems
Nathalie Majcherczyk, Nishan Srishankar*, Carlo Pinciroli
ICRA, 2021

Abstract

We implement data-driven realization for federated learning in a swarm system without the need for a central server through the used of a gossip-based conflict-free replicated data structure (CRDT).

Artificial Intelligence for the Advancement of Lunar and Planetary Science and Exploration
Indhu Varatharajan*, Valentin Bickel*, Daniel Angerhausen*, Eleni Antoniadou*, Shashwat Shukla*, Abhisek Maiti*, Ross Potter*, Nishan Srishankar*, Frank Soboczenski*, Carl Shneider*, Michelle Faragalli*, Mario D’Amore*
Planetary Science and Astrobiology Decadal Survey, 2020

Abstract

We explore AI-driven approaches and the need for open-sourcing models and datasets to reduce manual labor and aid future lunar/ planetary exploration through real-time data analysis.

Multi-Manual Grasping and Interaction in Real-Time Dynamic Simulations using a Penalty Based Approach
Adnan Munawar, Nishan Srishankar, Loris Fichera, Gregory Fischer
ICRA, 2020

Abstract | Video

We present a novel parametric method for real-time, realistic, multi-manual grasping and interaction with complex objects.

An Open-Source Framework for Rapid Development of Interactive Soft-Body Simulations for Real-Time Training
Adnan Munawar, Nishan Srishankar, Gregory Fischer
ICRA, 2020

Abstract | Video

We propose a framework which allows simulation of any generic, user-specified soft-body and interaction with a variety of input interface devices.

Projects
image

Udacity's Self-Driving Car Nanodegree

Abstract | Video
Ongoing.Completed Term One (Computer Vision & Deep Learning) and Term Two (Sensor Fusion, Localization, & Control).

Worked on multiple projects with a goal of implementation on Udacity's Carla (and possibly an RC-car). Implemented lane detection using color palette selection, Canny Edge detection, Region-of-interest determination, and Hough feature transform or color-edge thresholding, perspective transformation, line-fitting using sliding-window algorithm of a color-histogram and then second-order polynomial fitting, traffic sign classification based on custom neural networks (inspired by Lenet, Inception) on augmented German Traffic Sign Dataset, Behavioral Cloning using NVIDIA's end-to-end network, vehicle detection and tracking on a dashboard camera using generated features and SVMs (improved by then implementing ensembles and YOLOv2 for real-time tracking).

image

ICRA 2016 Challenge

Abstract | PDF | Website
Scaling Chain of Integrators Formal Methods challenge.

Created a control scheme to win the Scaling chain of integrators domain using ROS & Python. The problem required a point robot to optimally reach randomly-generated goals and avoid obstacles in any dimension-space. The solution used a traveling-salesman optimization to order goals, potential fields attracting to goals & repelling from obstacles, random walk to avoid unfortunately generated polytopes and logic to indicate unsolvable scenarios. A futher implementation was done by creating a bisimulation of the higher-order system with a lower-order system (which while less accurate will make finding a solution simpler.)

image

Artificial Intelligence

Abstract | PDF | Video
Deep Reinforcement Learning implementation for a racing game.

Implemented Deep-Deterministic Policy Gradient to control a racing agent in an Outrun simulator with rewards given based on normalized velocity and penalties given to collisions and off-road travel. The issue with this simulator is stochasticity (the agent can chose between different tracks at different points in the race) which meant significantly more training needed to be done, and the lack of proper sensors on the car (for positioning, proximity, wheel spin, orientation etc.) which meant that rewards couldn't be tuned as finely. This algorithm will be reimplemented on another simulator such as Torcs.

image

Multi-Robot Systems

Abstract | PDF | Video
Hardware implementations of collective motion and flocking in a Sparki swarm.

Worked on a setup using Sparki robots identified using Aruco tags and an overhead camera system. Implemented collective transport of of a bulky object, warehouse sorting (sorting of boxes into respective goal regions), and flocking. Flocking was implemented by creating a virtual distance sensor (using the overhead system) that would give local distances instead of a global distance. Each agent has feels a force (based on the Lennard-Jones potential relative to neighboring positions) and aims to maintain a similar velocity vector with other robots to a goal location.

image

Robot Dynamics

Abstract | PDF | Video
Haptic Feedback implemenation for the da Vinci surgical system.

Creation of a modular directional haptic feedback device that can augment the performance of a surgeon using the da Vinci research kit by providing sensory feedback in addition to visual feedback. The haptic feedback worked by using asymmetric sinusoidal ascillations producing an apparent shearing sensation implemented using resonating actuators using PSM joint commands in a ROS simulation environment. The goal was to make tasks like suturing much easier. An unanticipated side-effect of the setup was while directionality could originally be sensed, after a few minutes of constant vibration, the brain tunes out the sense of direction.

image

Compressible Fluid Mechanics

Abstract | PDF
Design and implementation of a Transonic wind tunnel.

Creation of a transonic wind tunnel to test air speeds betweem M=0.3 to M=0.9 and involved sizing calculations for the throat, Solidworks modeling of the system, with proper tolerancing before machining could occur.

Teaching
WPI_Seal

Artificial Intelligence Grader - Fall 2018
Instructor: Professor Dmitry Korkin

Introduction to Communication & Networks Graduate Tutor - Spring 2018
Instructor: Professor Alexander Wyglinski

Analysis of Probabilistic Signals and Systems Graduate Tutor - Fall 2017
Instructor: Professor James Matthews

Principles of Communication Systems Senior Tutor - Fall 2017
Instructor: Professor Alexander Wyglinski

Optimal Control Grader - Spring 2017
Instructor: Professor Raghvendra Cowlagi

Awards & Services
  • Reviewer:CDC17,AAMAS{19,20},NeurIPS{20,21,22},AAAI21,IEEE-RAL{20,21,22},ICLR{21,22},IROS{22,23}
  • ICRA 2016 FMR Scaling Chain of Integrators Winning team
  • WPI International Scholarship & Dean's List
  • Charles O. Thompson Scholar
  • Tau Beta Pi (Engineering Honor Society)
  • Edexcel Challenge Trophy & Best Academic Results 2011

Template: Courtesy of Jon Barron.