Projects

Research Internship @ Autodesk AI and Robotics Lab

• CAD Informed Adaptive Robotic Assembly; Developed an Assembly Motion Planner Algorithm; Python plugin for Fusion 360
• Integrated a concave-concave mesh collision library (C++) with In-house Robotic Environment; Impacted three research projects
• Initiated a deep learning project on Representation Learning of Assembly Motion Plans; Developed a representation schema for Motion Plans; Trained a VAE on a synthetic dataset; Applications in Grasp Authoring, Qualitative Assessment of Motion Plans
• 5 minute Showcase Presentation

Research Projects @ CAD Innovation and Engineering Lab

• Machine Learning

  • Representation Learning and Segmentation of mechanical motions; Trained motion classifiers for Type-Synthesis.
  • Developing Generative Models (Conditional GAN and VAE) for computational creativity and managing uncertainty.
  • Interactive ML Agent for Assisted Mechanism Design; Human Machine Collaboration; ML for conceptual machine design.
  • Working on deep RL model for Mechanical Design; Developed an OpenAI-gym Environment for Mechanism Design Task.
  • Developing a Machine Learning Driven Web-Based App for Synthesis of Mechanical Linkages; Deployed on Google Cloud Platform; url: http://motiongenpro.appspot.com/; [Tensorflow, React, Redux, MongoDB]

• Optimization

  • Developed Lagrange Optimization routine for four-bar linkage synthesis; Reduces constrained optimization into a polynomial system of equations.
  • Solved the system by groebner basis method; implemented using GIAC npm package on node.js server.
  • Lead Author of an award winning publication for solving practical synthesis problems (doi: 10.1115/1.4037801)

Paper Implementations and Course Projects

• Deep Reinforcement Learning for Continuous Control Tasks

  (https://github.com/deshpandeshrinath/deepDGP)

  • Implemented Deep DPG algorithm to learn continuous control policies
  • Compatible with all OpenAI-Gym environments.
  • Implemented Hindsight Experience Replay for learning goal-oriented tasks with sparse binary rewards.
  • Demo1, Demo2, Medium Blog

• End-to-End Monocular Visual Odometry by Deep Learning

  (https://github.com/sladebot/deepvo)

  • Built a deep Recurrent Convolutional Neural Network for pose estimation of a car
  • CNN was derived from pretrained FlowNet2.0
  • Trained and tested on KITTI visual odometry dataset (grayscale)
  • Supported by Human Interaction Lab, Stony Brook.
  • Demo

• Computing Consensus Trajectory

  • Developed an algorithm to find valid representative trajectory among n time stamped trajectories; works for any dimensional space.
  • Algorithm builds a weighted DAG on input; designed heuristics for assigning weights.
  • Consensus Trajectory is the dijkstra’s shortest path on DAG.

• Motion Planning and Optimal Control of a Car in Drifting Conditions

  • Designed Extended Kalman Filter for Observer; Modeled governing dynamics; Used empirical tire friction model for drift simulations.
  • Computed shortest path using Dynamic Programming. Obtained Optimal Control via Direct Collocation; Implemented in MATLAB using optimal control solver GPOPS II.
  • Used high gain PID controller to follow optimal control. Results match with empirical drifting techniques used by race drivers.
  • Report

• Motion Planning of Baxter Arm for Planar Pushing Task

  • Computed smooth B-Spline motion for pushing. Computed Jacobian matrix; Applied approximate Inverse Position Kinematics
  • Obtained joint angles and rates for the task. Performed simulations to validate the results.

• Interactive Manipulation of NURBS Surfaces

  • QT5, OpenGL based implementation in C++ for interactive manipulation of Non Uniform Rational B-Spline Surfaces.