Het Patel

Navigation · Perception · Robotics software

Experience

Company tenure: May 2026 – Present

Developing autonomous navigation using semantic SLAM for AutoMap, connecting local visual maps with facility graphs and cloud VLM environmental reasoning.
Deploying local navigation on Jetson Orin Nano and integrating cloud-assisted lost-localization recovery. Sub-10-second recovery is an engineering target, not a reported benchmark.

Jul 2024 – Aug 2025 · Ahmedabad, India

Developed and deployed ROS 2 navigation for construction-site material transport over GPS-denied routes longer than 200 m, connecting perception, localization, planning, and control.
Integrated LiDAR, IMU, wheel encoders, and RGB-D cameras through custom ROS 2 drivers; co-defined sensor placement and implemented online extrinsic calibration for realignment after on-site impacts.
Developed LiDAR–inertial localization and NDT alignment workflows with LIO-SAM and FAST-LIO2 for repetitive geometry in approximately 2,000 m² mapped environments.

May 2023 – Jan 2024 · Ahmedabad, India

Replaced a legacy recorder interface with a C++/Qt operator application for an indigenous SAR Solid State Recorder, using UART/RS-232 commands and USB 2.0 downloads.
Contributed to Gaganyaan cabin-display software using C++, OpenGL ES, and PetaLinux on a Xilinx platform.

Dec 2022 – Jun 2023 · Bangalore, India

Benchmarked YOLOv4, R-CNN, and BiT for smart-refrigerator food recognition; selected and deployed YOLOv4 with TensorRT in the device inference pipeline.

Master’s in Autonomy and Robotics · Completed

B.Tech in Electronics and Communication Engineering · Sep 2020 – May 2024 · GPA 3.8 / 4.0

Programming & systems: Python, C++, Linux, Qt, Git, Docker, CMake, Bash, pytest

Navigation & control: ROS / ROS 2, Nav2, Visual / LiDAR–inertial SLAM, Sensor integration, NDT alignment, MPPI, Stanley / PID, State machines

Perception & learning: OpenCV, PyTorch, CUDA, TensorRT, RGB-D / point clouds, SAM-3 / FoundationPose integration, SAC

Simulation & hardware: Isaac Sim, MuJoCo, Gazebo, Jetson Orin Nano, UR5e, UART / RS-232, USB, PetaLinux

December 2025 · Four-person team · Supervised tests on a physical Polaris GEM

Engineered a ROS 2 multimodal perception pipeline integrating voxel-filtered LiDAR DBSCAN clustering, EMA-based tracking, and YOLOv11 RGB-D localization; implemented 100-ms temporal synchronization, Euclidean data association, weighted sensor fusion, and discrete-time Time-to-Collision (TTC) estimation using predicted pedestrian and ego-vehicle trajectories.
Architected a hierarchical, risk-aware control framework integrating TTC thresholds, pedestrian motion classification, sensor-staleness detection, and regulatory-sign context; coupled Stanley lateral control with PID-based longitudinal regulation, state-dependent velocity mapping, and PACMod2 emergency-braking overrides for predictive stop/yield decisions.
Developed an integrated perception-to-actuation autonomy stack on the UIUC GEM vehicle, addressing the fundamental limitations of reactive pedestrian avoidance through short-horizon motion prediction and anticipatory collision-risk assessment, enabling context-aware speed adaptation, conservative handling of uncertain observations, and supervised vehicle operation.

December 2025 · Four-person computer vision project · Offline YCB-Video evaluation

Architected a language-conditioned 6D pose estimation framework integrating SAM-3's text-prompted instance segmentation with NVIDIA FoundationPose's render-and-compare architecture, enabling geometric pose registration, hypothesis scoring, and iterative SE(3) refinement for previously unseen objects under partial occlusion.
Engineered a modular perception and tracking architecture coupling Moondream2 semantic scene analysis, Objaverse-XL geometric proxy retrieval, and cached mesh management with dynamic target switching. Integrated dependency-isolated PyTorch/CUDA inference pipelines and characterized the SAM-3 latency bottleneck of approximately 12.5 seconds per frame.
Bridged natural-language object selection with manipulation-oriented spatial reasoning, transforming semantic descriptions into object-specific segmentation masks and geometrically grounded 6-DoF poses without requiring instance-specific model training. Evaluated on YCB-Video, achieving 76.5–100% ADD-S AUC across two objects, demonstrating the potential for language-directed robotic perception in cluttered environments.

Spring 2026 · MuJoCo training · OpenVLA in simulation · UR5e demo with scripted approach

Developed a 6-DoF Soft Actor-Critic (SAC) controller operating on 48-dimensional proprioceptive, TCP pose-error, and force/torque-history observations; implemented a latching contact-force handoff, joint-constrained residual actions, and exponential smoothing for stable 100 Hz contact-rich manipulation.
Engineered a MuJoCo-to-ROS 2 sim-to-real control pipeline for a UR5e–Robotiq manipulator, training across 21.6 million environment steps with dynamics and sensor randomization; developed stereo epipolar triangulation, ChArUco extrinsic calibration, and constrained Pinocchio QP-IK for target localization and manipulation.
Addressed the precision gap between vision-language manipulation and contact-constrained assembly through hierarchical control that separates coarse approach planning from learned insertion corrections, enabling submillimeter USB-A seating in a real-robot demonstration using a scripted approach policy, with OpenVLA integration evaluated separately in simulation.

Additional prototypes and design studies: thehetpatel.com/#projects