I did some work.

Others: Motion Planning: TrajOpt, RRT and Variants, RealTime RRT*; Image Segmentation with GMM, Image De- noising using MRF;

Autonomous Grasp Inference and Execution using Baxter and KUKA lwr4 Robots

Designed an end-to-end grasping pipeline to grasp objects on a table autonomously.
Training data was collected in Gazebo simulation and tested in real world. ISRR 2017

Video Action recognition using Deep Learning

Implemented a Bi-Directional LSTM Model on VGG16 Net using Keras to classify actions in scenes.
Achieved a Mean Average Precision of 15.7 mAP compared to the State of the Art of 21.4 mAP.

Simultaneous Robot State Estimation and Object Tracking

Implemented an Extended Kalman Filter algorithm to estimate the object pose from noisy JointState.
Used a Gaussian Mixture Model to plan a trajectory for Baxter arm to the object for grasping.

Grasp Collision detection using Convolutional Neural Networks

Developed a CNN model to detect collisions btw robot and environment using PointClouds and JointState.
Model classifies collisions with an accuracy of 84.7% and is ~30% faster than FCL.

Object Detection and Segmentation in Point Cloud data using PointNet

Trained modified PointNet model on YCB object dataset and BigBird dataset.
Model runs at 24 fps on a NVIDIA GeForce 1060 GPU with an accuracy of 88.3%.

Estimating Depth from a single image using FCN Network

Implemented a modified FCN Net and trained it on NYU Depth Dataset and KITTI Dataset.
Model achieved a mean RMSE error of 0.294 on NYU Depth and 0.312 on KITTI Dataset.

Real-time Semantic Segmentation on Low-Power Android Devices

Developed a fast background subtraction for portrait video based on modified SegNet model.
Model achieved a mean IoU of 87.3% at 30 FPS on Google Pixel 2.

Extrinsic Calibration of Stereo Camera and Velodyne LiDAR

Developed a ROS package to automate calibration between Velodyne VLP- 16 and ZED stereo camera.
Reduced the mean point to point error by 72% compared to manual feature based calibration.