Swapneel Wagholikar

I am currently a Robotics Master's student at Worcester Polytechnic Institute graduating in May 2024. I have had the opportunity to be a part of multiple projects affiliated to VISLab robotics lab at WPI. I work at the intersection of Perception, Deep learning, and embedded systems.

At WPI, I have worked on a Directed research, where I am advised by Prof. Ziming Zhang for implementation of PointAttN: Tranformer Network for Point Cloud Completion which aims to pay attention to the relationships between points without dividing shapes into smaller regions. During the summer of 2023, I completed an internship at Void Robotics, where my work revolved around Sensor Fusion of Odometry data from ZED2 depth camera and RTK-GPS Sensor.

Currently, I am working in DEKA Research and Development as a Robotics Software Intern (Perception) where I am working on Sentry bot to create high quality depth maps through classical computer vision and Deep Learning. Along with that, I am fulfilling the role of Deep Learning Researcher at Findability Sciences where I am developing an LLM-based conversational interface for business users.

I am actively looking for Full-time opportunities in Robotics, Deep Learning, Embedded Systems and Software industry starting May/June 2024. I believe that diversity can bring up new perspectives and will spark brilliant ideas. I look forward to learning with experts in academia and industry. Please feel free to reach out to me at swagholikar@wpi.edu


Connect with me on LinkedIn.

Email             Resume             GitHub

Optimizing a NeLF (Neural Light Field) based novel view synthesis, with techniques such as model pruning and quantization, and deployed it on iPhone, demonstrating proficiency in on-device deep learning for real-time 3D novel view synthesis.
Guide: Prof. Bashima Islam

In this project, sensor fusion of LiDAR and camera was carried out on the Kitti dataset to obtain painted pointcloud.
LiDAR pointcloud was projected on the images. Pointcloud segmentation is carried out on the images and each corresponding point in the pointcloud is classified and painted. ICP registration trials using Open3d are carried out to merge the pointclouds to obtain 3D reconstruction of the scene. Finally, for each pointcloud, the Bird's eye view is also obtained.

Panoptic segmentation was done on SemanticKITTI dataset (3D LiDAR Point Cloud Data) which combines the outputs of semantic segmentation and instance segmentation for Deep Learning under Prof. Fabricio Murai.
In this project, Shared encoder-decoder backbone was used to combine these two tasks. RangeNet++ was used for semantic segmentation, mask R-CNN for instance segmentation and implemented in TensorFlow 2.0 with novel parameter free panoptic head.

In this project, I have planned path traversal for non-holonomic robots by state-of-the-art algorithms.
For global path planning, I have used the most recent algorithms like AIT* and BIT*. These are the latest variants of Informed RRT*.
For local path planning and path traversal, APF and MPC are used. Evaluation is done based on time-complexity and accuracy of optimal path detection+traversal.

These projects include Neural Networks Design from Scratch, Neural Machine Translation.
Also, use of BERT Embeddings, Higher Level APIs, Design of Optimizers from Scratch in Numpy.
Pruning & Quantization, Neural Architecture Search, Dynamic Network Inference.

Simultaneously reconstructed 3D scene Mapping and extracted camera pose Localization from given stereo camera correspondences using Non-Linear triangulation, Non-Linear PnP and Bundle Adjustment (BA) pipeline.

In this project, Detected edges in image using probability based boundary detection using K-means clustering of Oriented DoG (ODoG), Leung-Malik (LM) and Gabor Filter bank responses. Outperformed the results from classical canny and sobel filters.

In this project, I implemented research work presented by Zhengyou Zhang for Calibrating a Camera which is one of the hallmark papers in camera calibration. Estimated parameters of the camera like the focal length, distortion coefficients and principle point.

In this project, we Implemented discrete action space algorithms such as DQN, DQN-MR and DQN-PER on OpenAI Gym’s Third party environment, Highway-env. Compared training time and accuracy to discover DQN-PER has the best performance.

Designed and deployed Sliding Mode Controllers for trajectory tracking for micro UAVs, with an acceptable error range of 1% while countering environmental noise. Generated a fifth-order trajectory with an accuracy of 0.03% for path planning.

This project implements various search based planners (Dijsktra, A star, D star) and sampling-based planners(RRT, RRT star and Informed RRT star).The planners are implemented on a 2d map.

These projects include implementation of Epsilon-Greedy Algorithm, Dynamic Programming, Monte Carlo, Temporal Difference Learning, Model-based RL

In this project, I reconstructed informed RRT algorithm for path planning of continuum robots which are needle sized manipulators used in image-guided surgical procedures resulted in accurately generating optimal and feasible path in a dynamic environment..

Implemented YOLO pipeline from scratch to do object detection on street scene images.
MAP acheived : 0.43

Great thanks to Jon for his amazing website template!