Kanishk Navale

Led abstraction & development of the company's specialized large language model product designed for robot interaction. Fine-tuned foundation models using LoRA methodology blended with geodesic multi-head self-attention for robot & visual object interaction. Engineered vision transformer extracting feature pyramid using PyTorch-Lightning framework for depth image completion with robust object mask estimation to infer valid depth images of an object and bring in geometrical aware boundary depth completion.

Awarded 1,3 (Excellent) grade for the master thesis with a scientific publication in one of the outstanding IEEE journals. Designed object semantic equivalent 3D keypoint regressing head on the backbone based on geometrical manifolds, making training procedures faster by 200% (highlight of the thesis) compared to existing methodologies without depth cameras, which are often noisy with present-day Technology.

Contributed pixel-wise distributional loss to optimize the deep learning network. The implemented loss function led the DNN to yield better uni-modal predictions visually by inspecting the heatmap of tracked keypoint over the existing loss function. Formulated an optimization process to regress geometrically consistent visual keypoints on the dense feature space by engineering self-supervised losses, eliminating manual keypoint selection to create oriented bounding boxes.

Integrated a robot physics engineer from scratch, introducing hybrid motion force control to the NYU-Finger robot (a leg of a quadruped robot) with the in-house fabricated capacitive haptic sensor. Devised robust noise elimination methodology using filters and state estimators for fluently reading the haptic sensor for robot manipulation.

Critiqued three student teams of a total size of 35 in industrial automation research and handled six projects simultaneously. Proposed & implemented General Adversarial Imitation Learning (GAIL) to improve gait control for a humanoid robot using C++ based ROS & Tensorflow 2, decreasing onboard computation costs by 2s.

Commissioned 48 industrial robots, 2D/3D vision, machine tending & IoT systems as part of the design & site deployment team. The technical & commercial proposals pitched generated revenues more than the set yearly target. Remodelled the cycle time estimation model to reconcile system breakdown time using an LSTM-based deep learning neural network, improving proposal surety by an in-house metric by 3 points.

Improved robot path planning by 60% fewer episode lengths using DDPG agent with prioritized & hindsight experience replay with parametric exploration noise using Tensorflow 2 & PyTorch. Implemented transformers-based speech & language learning model for classification with model optimizations, earning top-2 position in the class. Designed information maximization variation encoder (Info-VAE) for high data dimensional decomposition to its latent state, securing top-1 the class.

Led student team of size 10 to participate in ABU-Robocon international event engineering two badminton-playing robots. Built a delta-style parallel robot for the bachelor's capstone project. Designed mechanical robot system components using the SolidWorks CAD tool & inspected the manufactured parts based on the GD&T standards. Furthermore, based on control design, custom PCBs were fabricated using the EAGLE tool, and a PLC-based motion system was deployed using Beckhoff TwinCAT.