pROJECTS

I developed an imitation learning based control model to track feasible trajectories as closely as possible. I created a dataset with inputs and end effector positions to learn the reachability and dynamics of the space and then used imitation learning to learn a control policy that could reach arbitrary locations by providing correct inputs. The soft arm model was developed in MuJoCo and a PPO actor was learnt using StableBaselines3.

I worked on developing advanced control strategies for a Unitree B1 quadruped and a 6 DOF Unitree Z1 arm to mark areas in construction sites with exceptional precision, achieving accuracy up to 3 mm even on tilted surfaces. Our approach utilized RGBD point cloud data for robust navigation and manipulation capabilities, enabling seamless operation in cluttered environments. To streamline the system's functionality, we integrated the entire setup into a behavior tree interface with robust exception handling to switch and execute tasks as needed. Additionally, we designed and fabricated a two-link manipulator, which we mounted on a modified Unitree Go1 quadruped robot for precise re-bar marking via ink-jet technology, accurate to within 0.5 mm. We also developed custom kinematic controls to draw markers on specific waypoints, enhancing the versatility and efficiency of our solution.

Made this project in the course Robotic Caregivers, instructed by Dr. Chalrlie Kemp.
The main aim was to develop a robotic caregiving system "Blue" for older patients suffering from MCI or declining cognitive health in general. The system helps the patients to find household objects which they misplace throughout the day and guides them to the location of the objects.  An 'inventory' of the objects is maintained by patrolling the house several times a day and scanning common surfaces on which objects could be put on like kitchen counters, tables etc. It interacts with the user by showing the images of the most recent images of the objects it took and navigates around the house to show the user where the robot found the object. Check out the video for the demo!!

Biomimetics has been studied for a long time, and agile locomotion and manipulation is one area where learning from nature can be used to produce optimal locomotion strategies to traverse through difficult environments. In this project aims to establish a framework that allows learning agile movements from real animals that could be transferred to legged bipedal robots. A reinforcement learning-based pipeline with careful reward shaping can be established to learn a diverse class of behaviors that are agile and can not be learned by training policies traditionally. Future work involves integrating a manipulator on the existing bipedal Cassie robot and moving to a combination of RL and manipulator controls.

Preliminary Report - Report

Traditional linear control strategies have been extensively researched and utilized in many robotic and industrial applications and yet they don’t respond to the total dynamics of the systems. To avoid tedious calculations for nonlinear control schemes like H-infinity control and predictive control, the application of Reinforcement Learning(RL) can provide alternative solutions. This article presents the implementation of RL control with Deep Deterministic Policy Gradient and Proximal Policy Optimization on a mobile self-balancing Extendable Wheeled Inverted Pendulum (E-WIP) system. Such RL models make the task of finding satisfactory control schemes easier and responding to the dynamics effectively while self-tuning the parameters to provide better control. In this article, two RL-based controllers are pitted against an MPC controller to evaluate the performance on the basis of state variables of the E-WIP system while following a specific desired trajectory.

Autoencoders and generative neural network models have recently gained popularity in fluid mechanics due to their spontaneity and low processing time instead of high fidelity CFD simulations. Autoencoders are used as model order reduction tools in applications of fluid mechanics by compressing input high-dimensional data using an encoder to map the input space into a lower-dimensional latent space. Whereas, generative models such as Variational Auto-encoders (VAEs) and Generative Adversarial Networks (GANs) are proving to be effective in generating solutions to chaotic models with high ‘randomness’ such as turbulent flows. In this study, forced isotropic turbulence flow is generated by parameterizing into some basic statistical characteristics. The models trained on pre- simulated data from dependencies on these characteristics and the flow generation is then affected by varying these parameters. The latent vectors pushed along the generator models like the decoders and generators contain independent entries which can be used to create different outputs with similar properties. The use of neural network-based architecture removes the need for dependency on the classical mesh-based Navier-Stoke equation estimation which is prominent in many CFD softwares.

Many unconventional descent mechanisms are evolved in nature to maximize the dispersion of seeds to increase the population of floral species. The induced autorotation produces lift through asymmetrical weight distribution, increasing the fall duration and giving the seed extra time to get drifted away by the wind. The proposed bio-inspired concept was used to produce novel modern pods for various aerospace applications that require free-falling or controlled velocity descent in planetary or interplanetary missions without relying on traditional techniques such as propulsion-based descent and the use of parachutes. We provide an explanation for the design procedure and the functioning of a mono blade auto-rotating wing. An element-based computational method based on Glauert's blade element momentum theory (BEMT) model was employed to estimate the geometry by maximizing the coefficient of power through MATLAB's optimization toolbox using the Sequential quadratic programming (SQP) solver. The dynamic model was developed for the single-wing design through the MATLAB Simulink 6-DOF toolbox to carry out a free-flight simulation of the wing to verify its global stability.

As a part of the Cansat 2020 Competition, this project was aimed at developing a  deployable delta-wing glider that is launched at high altitude and glides down in a specific trajectory autonomously using onboard control surfaces and sensors. This project required the designing of a gilder body which encloses the deployment mechanisms, a deployable parachute, foldable wings, a gyro stabilized camera, onboard sensors, communication systems and power systems. This glider had to sense the air pressure, temperature, wind speed and air particulate levels while gliding down in a specified spiral trajectory. Such systems make for great environmental survey devices and can be used in rescue operations as well.