BIOGRAPHY

Hi, I’m Mukul Ganwal - a passionate roboticist driven by my love for robots and their potential to transform our world.

Here’s my Curriculum Vitae for detailed information.

I am a Senior Perception Engineer and Team Lead at Anyware Robotics, where I rebuilt and grew the perception team to 5 engineers and own the roadmap across 3D perception, camera systems, and learning-based work. My work focuses on building production-grade machine learning pipelines that enable robots to understand and interact with complex real-world environments. I ship systems that use three RGB-D cameras to reconstruct work cells in 0.1 seconds, allowing FANUC CRX25 robots to palletize boxes in 6.3 seconds each at 573 boxes/hour.

Most recently, I architected a custom dual-head, NMS-free segmentation model end-to-end — adding a wireframe pose-detection head that enables direct 3D box pose estimation — reaching 99.4 mAP@50 while cutting inference from 56 ms to 7 ms. I also overhauled the 2D/3D perception pipeline, parallelizing preprocessing across three cameras and refactoring the point-cloud stack to cut 2D compute from 900 ms to 90 ms and 3D from 2 s to 300 ms; hiding perception behind arm motion brought the pick cycle from 12 s down to 9 s and lifted throughput by 100+ cases/hour. I’m also a co-inventor on a U.S. provisional patent for multi-purpose robotic systems for warehouse automation.

My expertise lies at the intersection of robotics, computer vision, and MLOps. I specialize in creating active-learning workflows that dramatically reduce data labeling requirements - achieving 40% reductions in manual annotation while maintaining rapid 40-minute test-to-deploy cycles. I’ve pioneered the use of synthetic data generation through Blender photogrammetry combined with StyleGAN3 domain randomization, scaling datasets by 8× and improving detection performance by 4.7 percentage points. My systems integrate seamlessly with ROS 2 architectures, incorporating behavior trees and hierarchical finite-state machines with real-time fault classification that has reduced unplanned robot downtime by 35%.

Before joining Anyware Robotics, I honed my skills as a Graduate Research Assistant at the BioRobotics Lab in Pennsylvania, where I engineered modular battery extraction mechanisms for Apple devices and built real-time multi-class instance segmentation pipelines achieving 98% accuracy. My journey into robotics and AI began with internships at PGP Glass and other companies, where I developed visual inspection systems with 99% defect detection accuracy and created intelligence systems that transformed business operations.

My technical toolkit spans PyTorch, JAX, ONNX, TensorRT, CUDA, OpenCV, Open3D, PCL, and ROS/ROS2 for ML, perception, and robotics, complemented by Docker, Kubernetes, Ray, Git, AWS, and GCP for robust MLOps deployments. I’m proficient in Python, C/C++, MATLAB, and Bash across Ubuntu LTS/Pro RT and FreeRTOS environments, with expertise in simulation platforms like Isaac Sim, Gazebo, and PyBullet alongside CAD tools like SolidWorks.

What drives me is the belief that robots can fundamentally improve how we work and live. Every system I build bridges the gap between cutting-edge AI research and practical robotic applications that deliver measurable impact in real manufacturing environments.

“My heart in my work” - Andrew Carnegie