Investigated several methods to find an optimal or near-optimal grasp on convex shapes that produces force/form closure for both planar and spatial objects and also evaluated the grasp quality. [Project Report]
Set up an end to end system with state of art dense mapping algorithm to improve object surface reconstruction for performing accurate grasp synthesis. This entailed performing dense 3D reconstruction, point cloud segmentation of objects, filtering (MLS and Poisson surface generation) and using the final model for force closure and manipulation evaluation. [Project Report]
Using 3D feature histograms on dense scans to train models for food bite locations. Using these trained classifiers on unseen food scans to do detection in clutter and predicted bite location using supervised machine learning models with an accuracy of 96%. [Project Report]
Integrated inertial data (IMU) to inform optimization to reduce drift in dense mapping algorithm (ElasticFusion) to create consistent global maps. [Project Report]
Analyzed and implemented Generalized iterative closest point (G-ICP) algorithm and compared it with standard Iterative Closest Point algorithm for doing 3D point cloud registration pairwise and across multiple frames, done for 1000+ kinect frames. [Project Report]
Creating dense 3D maps for large surfaces with millimeter level accuracy.
- Dense mapping: This entailed stereo camera disparity matching, camera calibration and dense depth frame generation. Mapping and localization using RGB frame tracking and depth frame matching via point-to-plane ICP.
- 3D Point Cloud recognition using Machine Learning: Represented 3D point cloud based object features using Point feature histograms (PFH and its variants FPFH). Trained non-linear classifier (SVM and RFC) for various 3D feature identification. Used Conditional random fields (CRF) for filtering in order to achieve accurate scene recognition to apply to surface damage detection and assessment.
- Constraint based state integration: Used sensor fusion to integrate inertial state to improve loop closure and reduce drift. Applied pose graph based constrained optimization for accurate mapping and localization (in progress).
NASA Ames Research Center, NASA, Moffett Field, CA, USA
- Wind turbine inspection using helicopter UAVs for GE R&D Labs. Patent filed by GE:US 20120136630.
- Co-PI for development of autonomous helicopter MAV (under 500gms) with full sensor telemetry (wireless), ground software & instrumentation for MSRSAS, Bangalore.
- Led the team in designing full indigenous Inertial measuring unit (IMU) quaternion based Extended Kalman Filter 9 DOF (3-axis Acc/Gyro/Mag), with no “Gimbal Lock” issues and attained a calculation response rate of over 300 hz.
- Built fully indigenous Autopilot using IMU and applied detailed control system theory and PID loops for stabilization of the UAV.
- Designed and built Helicopter UAV systems of 1 kg, 3 kgs, 12 kgs and 20 kgs AUW with indigenous chassis designs.
- Created patentable designs of robots which carry X-Ray emitter and detector for external oil/gas pipeline inspection.
- As a consultant for Team Indus, Google Lunar X-prize participant, authored report on lunar conditions and created detailed specification for a lunar rover.
- Working with students in UC Berkeley Lab, integrated Mavlink ROS node to Ardupilot and developed external range PID loop and Kalman Filter based ranging control system for constant distance flight from wall surface.
- As a student in Singularity University, built a prototype design for 3D printed robotic oral teeth cleaning solution. US Patent Publication no. US20160206415 A1.
- Developed embedded system control for multiple large volume enterprise storage system. This required controlling logical volume manager and adding virtualization features to enable RAID-1 to 5.
- Awarded two patents for the research on building distributed storage architecture over Logical volume manager (LVM) to repurpose unused drive space with iSCSI node and create one large virtual storage with node redundancies.