Projects
Real-time Planning for Robot Self-defense
Developed a real-time motion planning algorithm for a robot self-defense scenario, where the robot is tasked with protecting itself from incoming objects.
Deployed the planner on the PR2 robot, performed integration with the vision pipeline, and provided live demonstrations to project sponsors.
The motion planning stack that was developed is currently being run by Lockheed Martin and the US Army Research Lab for their applications.
Related Publications -
Blink and You'll Miss: A Planning and Perception Framework for Intercepting Projectiles Using Robot Manipulators; Yash Oza, Manash Pratim Das, Fahad Islam, Muhammad Suhail Saleem, Maxim Likhachev (In submission).
Real-time planning and perception for PR2 robot protecting itself from an incoming object
Motion Planning for Articulated Objects
Developed a Learning from Demonstration-based motion planning system for manipulating articulated objects using a 7-Dof robot arm.
Proposed a novel approach to speed up the existing algorithm by learning useful macro-actions offline and using them during online execution.
Deployed the framework on the RoMaN robot (designed by NASA), and provided live demonstrations to top US government officials for the RCTA capstone event.
Media coverage: The Economist, The Robot Report
Related Publications -
Human-Scale Mobile Manipulation Using RoMan, Journal of Field Robotics 2021; Journal paper with collaborators from NASA JPL, U.S. Army Research Laboratory, Univ. of Pennsylvania, Univ. of Washington, Lead author from CMU (in print)
Fast and High-Quality GPU-based Deliberative Object Pose Estimation, Journal of Field Robotics 2021; Aditya Agarwal, Yash Oza, Maxim Likhachev, Chad C. Kessens (in print)
Computing Probably Approximately Correct Set of Macro-actions for Heuristic Search-Based Planning, In submission; Yash Oza, Dhruv Mauria Saxena, Manash Pratim Das, Maxim Likhachev
The RoMaN robot performing a constrained manipulation task of opening a container
Planning and Execution using Inaccurate Models
Developed a motion planning framework that makes use of inaccurate dynamical model in order to complete a goal-oriented robotic task.
The algorithm ensures that the robot provably reaches the goal online without any access to resets.
Related Publications -
Planning and Execution using Inaccurate Models with Provable Guarantees, 2020 Robotics: Science and Systems; Anirudh Vemula, Yash Oza, J Andrew Bagnell, Maxim Likhachev
PR2 robot successfully performing a pick-and-place task with a faulty dynamics model
Model Predictive Control for Autonomous Driving
Developed a novel hierarchical optimization algorithm to compute trajectories for a self-driving car navigating among multiple dynamic obstacles.
The optimization problem computes a path for the vehicle followed by a set of velocities along the path, such that the dynamics of the vehicle and the collision avoidance constraints are satisfied while minimizing a user-defined cost function.
The algorithm was tested on challenging lane-changing, overtaking and lane-merging scenarios in the Gazebo simulator.
Related Publications -
Model predictive control for autonomous driving based on time scaled collision cone, 2018 ECC; Mithun Babu*, Yash Oza*, Arun Kumar Singh, K Madhava Krishna, Shanti Medasani (*denotes equal contribution)
Overtaking maneuver performed by the vehicle (in grey) using our MPC algorithm
Two-wheeled self-balancing robot
Performed mathematical modeling of a two-wheeled Self-Balancing robot, using Lagrangian mechanics.
Designed and implemented the PID, LQR, and Pole-Placement-based controllers.
Demonstrated position and velocity tracking using state integral feedback
Our two-wheeled self-balancing robot balancing on a sloped surface using a LQR controller