Parv Maheshwari

Sep '23

May '23

Jul '22

May '22

Apr '22

Feb '22

Sep '21

Sep '21

May '21

May '21

Apr '21

Apr '21

Mar '20

Working under the supervision of Prof. Sebastian Scherer and Dr. Wenshan Wang at Airlab, Robotics Institute, CMU on developing generalisable and accurate vehicle models. My primary mission is to enhance the autonomous navigation capabilities of a modified Yamaha All-Terrain Vehicle by developing an accurate dynamics model which considers the effect of both the control input and the ATV's environment.

Worked under the supervision of Dr. Bharadwaj Amrutur and Dr. Mukunda Bharatheesha . Worked on a centralised task allocation module and motion planning of warehouse robots. Extended this work to a networked setting where robots could cimmunicate with each other to actively avoid colliding. This work was submitted at ICRA, 2022.

Working under the supervision of Dr. Debashish Chakravarty at AGV, IIT Kharagpur in the motion planning team as a team meber in my inital years. Here I got to worlk on local planners. In my junior year , I led the team and our performance was tremendous in IROS Navigation and MLRC challenge.

The task was to develop (in simulation) a mobile robot with a manipulator capable of multiple waypoint navigation and exploration. As the winning solution, we created a probabilistic route planner, which finds near-optimal solutions while respecting the computational capabilities of the robot and integrated this with an adaptive MPC.

We presented this work at the International Conference on Control and Robotics 2022 in our paper "Multiple Waypoint Navigation in Unknown Indoor Environments"

Member of the IIT Kharagpur - IUPUI, Indiana - USB Colombia collaborative team.

Co-organised by IIT Kanpur and Drona Aviation, as part of the Inter IIT Tech Meet 11.0. This challenge aimed to develop a vision based state feedback control for an indoor multi-drone system handling socket communication with the flight controller without the use of ROS. Being a core team menber, I played a pivotal role in leading the communication and motion planning team.

We used a adaptive PID controller which received the current pose estimate on basis of precise detection of Aruco Tag using our vision module. Our team secured the first place.

In spring 2023, my aim was to enhance dynamics modeling for out-of-distribution (OOD) scenarios while maintaining in-domain accuracy. I explored Physics-Informed Neural Networks (PINN), which merge accurtacy of learning based methods with the generalization capability of traditional physics models. PINNs yielded significant improvements in modeling vehicle dynamics compared to solely data-dependent neural networks. However, during the summer of 2023, I encountered a key limitation of PINN: it required well-spread data for physics evaluations, a challenge in off-road driving due to data collection complexities.

To address this limitation, I introduced "PIAug - Physics Informed Augmentation," currently under review for ICRA 2024. This approach involved using a physics priior for both augmenting a dataset while also introducting the phsyics principles to NN. I then optimised the deployment of the NN-based vehicle models by reducing the inference runtime by 6x and GPU memory requirements by 1.9x. As a usecase PIAug allowed the vehicle to navigate with 4x tighter tracking constraints at out-of-domain velocities.

During my first summer at CMU in 2022, I focused towards developing an online prediction system for a dynamics-based vehicle model. This approach involved harnessing information from the robot's history, such as odometry and IMU data, and incorporating additional inputs about the future's environemnt like FPV images and processed top-down terrain maps. As a result of these efforts, we achieved a noteworthy 25% improvement in modeling accuracy for the vehicle's dynamics. However, the approach revealed limitations in handling out-of-distribution (OOD) scenarios, potentially leading to unsafe predictions of parameters. Despite these setbacks, this project introduced me to the stimulating field of aiding robots in navigating challenging environments. While we couldn't deploy this method on the ATV, it strengthened my determination to continue working towards enhancing robotic navigation in demanding conditions.

The objective of this project was to develop a high-performance network infrastructure to create a safe network-aware autonomy for multiple robots. Being one of the earliest members in the team, I was responsible with developing the Multi Robot Task Allocation module. I did this using a market auction based approach which considers the time required to complete the already queded tasks of each robot before finding the robot which would optimize the overall time required by the fleet to execute the tasks.

After this, I developed a Multi Robot Path Planning module by extending the existing NAV2 library to support multirobot navigation. At last, I proposed and showcase how sharing ego-footprint information with other robots in vicinty over a shared network can help in proactively avoiding the other robots. This also allows robots to avoid collisions even when other robots cannot be detected using sensors due to occluded sensing range. Example - when 2 robots are coming from round the corner and are not detected by each other's lidar or camera due to obstruction in sensing. This work was submitted at ICRA 2022.

Led a team of 11 to develop an Open Source Python and C++ implementation for a sampling based planner in Frenet Frame with support for both ROS1 and ROS2 support along with simulation support of Gazebo and LGSVL. OpenMP was used for increasing the speed of the sampling and map updation step while planning. In its early days, this project acted as the first step towards parallelized computing and later this also became the first reasearch project where I led a team :)

Implemented an optimized tiled and batched Cholesky factorization in CUDA using only single-precision arithmetic. Extensive experiments were then conducted on advantages of chunking while tuning the parameters of tile size, and order of looking. This project helped me get aquantied to CUDA and helped me understand how to efficiently exploit its use by efficient use of continuous threads and blocks.

Developed a primary healthcare disease checker bot as a combination of a traditional multi-layer perceptron neural netowork and steps of evolutionary/genetic algorithms like – Selection, Crossover, and Mutation. Used SymCat, a public dataset mapping diseases to various symptoms, as the training and validatiaon dataset. Further compared this hybrid method against standard classifiers based on Naïve Bayes Classifier, neural networks and random forest.