Work Experience
WORK EXPERIENCE
April 2021 Onwards:
Robot Control Software Engineer at Saga Robotics Ltd working on motion planning, control and perception for autonomous strawberry harvesting
May 2019 - March 2021
Post-doc at Surrey Space Centre: Lead developer of the simulator Spacecraft Manipulator Arm Simulation Hybrid (SMASH) engine with capability for autonomous trajectory learning for planning and control of free-floating spacecraft arms for space debris removal, on-orbit servicing and rendezvous & docking.
1. Path planning of free floating spacecraft manipulator arm
March 2018 - March 2019
Post-doc at Lincoln Centre for Autonomous Systems: Learning from demonstrations, optimization, robot manipulation, end-effector design and control
I worked on robot learning from kinesthetic demonstrations and motion planning for manipulators. I used Franka Emika’s 7-DOF panda robot arm which was controlled through ROS and python scripting. For perception, I used a static Kinect 3D sensor for scene segmentation, mapping (octo-mapping) . For determining the target pose, an on-hand Intel realsense depth camera was used. The point cloud data is further used for 3D octomap generation for obstacle avoidance. I had worked with sampling-based motion planning and obstacle avoidance algorithms in MoveIt along with Rviz but as the scene becomes cluttered, the trajectory produced becomes clumsy. For a better solution, I had come up with the idea of combining learning from kinesthetic demonstration along with gradient based obstacle avoidance to obtain a human like motion plan. For this, the learning was carried out using Probabilistic Movement Primitives (ProMPs). The main advantage of ProMPs over optimal control was that it gives a distribution of feasible trajectories. The mean of this trajectory distribution would be the solution provided by optimal control. Another significant advantage was that the trajectory distribution learned using ProMPs could be conditioned in task or joint space to get several feasible trajectories to the same task space point which could be used as initialization for optimization routines. The optimization carried out in the parameter space would not only ensure that the cost of the trajectory in terms of smoothness and being near to the obstacles was minimal but also computationally inexpensive.
1. Improving Local Trajectory Optimization by using Probabilistic Movement Primitives, IROS 2019
2. Playing back learned trajectories on a simulated 7-DOF panda robot
3. Simulation of trajectory optimization to avoid obstacles
April 2017 - February 2018
Post-doc at Robotics and Design Lab: On-Board Passive Micro-Vibration Isolation in Spacecrafts
The focus of the project was on modeling, simulation and fabrication of a Stewart platform based micro-vibration isolator.
In a spacecraft there are several sources of micro-vibrations and these significantly degrade the performance of precision instruments and sensors on board. The aim of this project was to arrive at the design parameters which make the first six modal frequencies close to each other thus enabling effective vibration isolation for all six primary modes. The modeling was carried out in MATLAB. For our experiments, the base of the platform was excited with an acceleration sine sweep of constant amplitude of 0.5g sweeping from 5-100 Hz. The lateral modal survey test was conducted while the platform was mounted on the slip table and with same excitation signal used for longitudinal test. The response was measured using accelerometers mounted on the center of bottom and top platforms. Our team could achieve a dynamic isotropy index (omega_max/omega_min) of 1.5 for our first prototype.