I am currently working as Autonomy Engineer at Caterpillar, where I develop planning algorithms for heavy construction & mining machineries. I work at the intersection of Motion Planning, Controls and Machine Learning solving problems in engineering and research in robotic navigation be it for mobile, legged or manipulator systems. In my previous work experience, I was involved in the development of multi-zone cleaning & AutoMode features for autonomous vacuum cleaners. I was involved through the stages of design, implementation, testing & deployment of the features.
I graduated in May 2021 with a Masters Degree in Robotics Engineering from Worcester Polytechnic Institute. At WPI, I was advised by Prof. Jing Xiao for extending genetic algorithm based motion planning framework RAMP for social robot navigation by integrating the Social Force Model into cost function optimization using Reinforcement Learning.
Prior to my graduate studies, I spent a year as a Visiting Research Scholar in Biorobotics Lab at Carnegie Mellon University working under the guidance of LTC Steve Crews, Dr. Matthew Travers and Prof. Howie Choset.
My Undergraduate Thesis with Dr. Shital Chiddarwar (VNIT Nagpur) involved development of Multi-Snake Modular Robot system to obtain legged robot traverse capability from combination of snake robots. In addition to work, I also mentor a group of students at IvLabs, where I conduct tutorials and mentor projects on Motion Planning and Controls for Underactuated Robotics Systems.
For any queries, please feel free to reach out to me at ssagrawal@wpi.edu.
I graduated in May 2021 with a Masters Degree in Robotics Engineering from Worcester Polytechnic Institute. At WPI, I was advised by Prof. Jing Xiao for extending genetic algorithm based motion planning framework RAMP for social robot navigation by integrating the Social Force Model into cost function optimization using Reinforcement Learning.
Prior to my graduate studies, I spent a year as a Visiting Research Scholar in Biorobotics Lab at Carnegie Mellon University working under the guidance of LTC Steve Crews, Dr. Matthew Travers and Prof. Howie Choset.
My Undergraduate Thesis with Dr. Shital Chiddarwar (VNIT Nagpur) involved development of Multi-Snake Modular Robot system to obtain legged robot traverse capability from combination of snake robots. In addition to work, I also mentor a group of students at IvLabs, where I conduct tutorials and mentor projects on Motion Planning and Controls for Underactuated Robotics Systems.
For any queries, please feel free to reach out to me at ssagrawal@wpi.edu.
Topics of Interest
Controls: Trajectory Optimization, Model Predictive Control, Optimal Control, System kinematics and dynamics
Planning: Sampling based Planning Algorithms, Motion Planning, Route Planning, Trajectory Optimization, Data-driven and Model based combined Planning
Machine Learning: Seq2Seq Prediction, Generative Models (GANs), Reinforcement Learning(RL), Imitation Learning (Inverse RL)
Controls: Trajectory Optimization, Model Predictive Control, Optimal Control, System kinematics and dynamics
Planning: Sampling based Planning Algorithms, Motion Planning, Route Planning, Trajectory Optimization, Data-driven and Model based combined Planning
Machine Learning: Seq2Seq Prediction, Generative Models (GANs), Reinforcement Learning(RL), Imitation Learning (Inverse RL)
Experience
Robotics Software Engineer Managed by Mr. Sarath Suvarna
San Jose, CA (June 2021 - Nov 2022)
Feature Lead:
Responsible for R&D of Automode feature that enables robot to identify the surface type & switch between cleaning modes to optimize for cleaning & power performance. I was managing a team of 2 SWE & 2 Embedded QA
members.
Features Team: Responsible for improving the robot navigation for multi-zone cleaning feature.
Navigation Team: Responsible for resolving any navigation failures due to bad path planning. Added visualization features & improved simulator to aid debugging planning/mapping issues. Implemented an incremental search algorithm for avoiding dynamic obstacles in the environment. I contributed to the team in improving navigation stack to achieve 98.6% autonomy between docking & undocking.
San Jose, CA (June 2021 - Nov 2022)
Features Team: Responsible for improving the robot navigation for multi-zone cleaning feature.
Navigation Team: Responsible for resolving any navigation failures due to bad path planning. Added visualization features & improved simulator to aid debugging planning/mapping issues. Implemented an incremental search algorithm for avoiding dynamic obstacles in the environment. I contributed to the team in improving navigation stack to achieve 98.6% autonomy between docking & undocking.
Robotics Research Intern Supervised by Dr. Soshi Iba
San Jose, CA (Jan 2021 - May 2021)
San Jose, CA (Jan 2021 - May 2021)
Intention Estimation for robot teleoperation:
Given the sequence on robot motion during teleoperation, the task is to estimate the teleoperator's intended task and provide assistance in executing the task. To solve the problem, I developed an Inverse Reinforcement Learning based probablistic method that estimates the intended tasks of the teleoperator in shared autonomy and evaluated the proposed algorithm in custom proof-of-concept gym enviornment.
Adv. Robotics Engineer managed by Mr. Felipe Polido
North Reading, MA (Aug 2020 - Dec 2020)
North Reading, MA (Aug 2020 - Dec 2020)
Systems Engineering: Worked as a Systems Engineer with the Manipulation team, integrating and testing perception and planning modules in tote consolidation pipeline across hardware and simulation platforms. I was also involved in setting up the manipulation station from ground up and calibrating sensors.
JVSRP Intern supervised by: Mr. Rohan Thakker, Dr. Hiro Ono and Mr. Kalind Carpenter
Pasedena, CA (May 2020 - Aug 2020)
(Image courtesy by Keck Institue for Space Studies - Caltech)
Given the robot's kinematic model, contact locations and state-action constraints, find the optimal contact force that the robot should apply on the crevasse to enable stable motion against the plume forces. To solve this problem, I worked with Robotic Aerial Mobility (347T): Exobiology Extant Life Surveyor (EELS) software team in developing grasping based new planning & control framework for snake robot climbing in crevasse for Enceladus exploration [EELS Concept Video].
Pasedena, CA (May 2020 - Aug 2020)
Given the robot's kinematic model, contact locations and state-action constraints, find the optimal contact force that the robot should apply on the crevasse to enable stable motion against the plume forces. To solve this problem, I worked with Robotic Aerial Mobility (347T): Exobiology Extant Life Surveyor (EELS) software team in developing grasping based new planning & control framework for snake robot climbing in crevasse for Enceladus exploration [EELS Concept Video].
Graduate Research Assistant supervised by: Prof. Jing Xiao
Worcester, MA (Aug 2020 - May 2021)
(left: Dynamic obstacle detection, center and right: navigation using RAMP with simulated pedestrians)
Social Robot Navigation: Real-time Adaptive Motion Planning (RAMP) is genetic algorithm based framework for planning and execution of non-holonomic robot trajectories in an environment with dynamic and obstacles. The cost function used for evaluating the trajectories in the population is a weighted combination of multiple optimization criteria such as time, energy, safety and social behavior. However a single set of parameters in the Social Force Model(SFM) fails to adapt to the changing and unseen environment. To tackle this problem, we use LSTM-DQN to self-adapt the SFM parameters in social robot navigation. For training purpose, we simulate the pedestrian motion using the pedsim environment.
Worcester, MA (Aug 2020 - May 2021)
Social Robot Navigation: Real-time Adaptive Motion Planning (RAMP) is genetic algorithm based framework for planning and execution of non-holonomic robot trajectories in an environment with dynamic and obstacles. The cost function used for evaluating the trajectories in the population is a weighted combination of multiple optimization criteria such as time, energy, safety and social behavior. However a single set of parameters in the Social Force Model(SFM) fails to adapt to the changing and unseen environment. To tackle this problem, we use LSTM-DQN to self-adapt the SFM parameters in social robot navigation. For training purpose, we simulate the pedestrian motion using the pedsim environment.
Visiting Research Scholar supervised by: LTC Steve Crews, Dr. Matthew Travers and Prof. Howie Choset
Pittsburgh, PA (June 2018 - July 2019)
(Video courtesy by LTC Steve Crews and Abhimanyu Singh)
Unified Foothold Selection and Motion Planning for Legged Systems in Real-Time: In scenerios, where robot is subjected to large disturbances, conventional frameworks first re-plan the desired footstep location, then plans motion that moves the body forward and finally derives a controllers that stabilizes the system about the planned motion. However, in practice these individual steps are all coupled. In this work we present a novel architecture that unifies footstep planning, motion planning, and online feedback control for legged robots moving through complex environments. To embed the footstep selection into the planning and control framework, we use the terrain height map as the soft constraint in the MPC-iLQR optimization. Beside this, the analytical linearization of hybrid dynamics allows computation speed of upto 7x as compared to available numerical differentiation methods. Allowing the robot to leverage its underactuated dynamics in avoiding sudden and unforeseen obstacles in the environment [IEEE Humanoids 2019 Paper].
Pittsburgh, PA (June 2018 - July 2019)
Unified Foothold Selection and Motion Planning for Legged Systems in Real-Time: In scenerios, where robot is subjected to large disturbances, conventional frameworks first re-plan the desired footstep location, then plans motion that moves the body forward and finally derives a controllers that stabilizes the system about the planned motion. However, in practice these individual steps are all coupled. In this work we present a novel architecture that unifies footstep planning, motion planning, and online feedback control for legged robots moving through complex environments. To embed the footstep selection into the planning and control framework, we use the terrain height map as the soft constraint in the MPC-iLQR optimization. Beside this, the analytical linearization of hybrid dynamics allows computation speed of upto 7x as compared to available numerical differentiation methods. Allowing the robot to leverage its underactuated dynamics in avoiding sudden and unforeseen obstacles in the environment [IEEE Humanoids 2019 Paper].