Robotics Engineer

Pranay Katyal

Worcester, MA +1-774-***-**** *************@************.*****.*** github.linkedin.io com/in/pranay-katyal github.com/pranaykatyal Professional Summary

Robotics Engineer with an MS in Robotics Engineering (WPI, May 2026, GPA 3.8) seeking full-time industry roles in robotics software, autonomy, and embedded systems. Hands-on experience building, debugging, and deploying real hardware systems including robotic arms, quadcopters, and autonomous ground vehicles using C++, Python, and ROS2. Proven ability to take systems from simulation through hardware commissioning, with a strong foundation in control systems, sensor integration, motion planning, and perception. Contributor to a peer-reviewed paper submitted to IEEE RA-L. F-1 visa; OPT EAD begins June 15, 2026.

Education

Worcester Polytechnic Institute Worcester, MA

MS, Robotics Engineering — GPA: 3.8/4.0 2024 – 2026 Coursework: Foundation of Robotics (A), Robot Dynamics (A), Robot Control (B), Motion Planning (A), Multi Robot Systems

(A), ML for Robotics (A), Deep Learning (A), Autonomous Aerial Vehicles (A), Safety Guarantees in Autonomous Robots (A), Directed Research (A), Capstone (A), Computer Vision (A) Chandigarh University Chandigarh, India

BS, Mechatronics Engineering — GPA: 3.4/4.0 2019 – 2023 Technical Skills

Programming: Python, C++, MATLAB, Bash/Shell Scripting, CMake Robotics & Control: ROS2 (Humble, Jazzy, Foxy), PID/LQR Control, CBF/GCBF Safety Filters, State Estimation

(UKF/EKF), Motion Planning (RRT*, OMPL), Trajectory Optimization, SLAM, Visual Servoing, Sensor Fusion Hardware & Sensors: Dynamixel Motors, IMU Sensors, SICK TIM781S LiDAR, Event Cameras (GenX320, Prophesee EVK4), CrazyFlie 2.0, AWS DeepRacer, OpenManipulatorX (4-DOF), Kinova Gen3 (6-DOF), Tello, ArduPilot, PLC (Ladder Logic) Perception & ML: PyTorch, TensorFlow, OpenCV, RAFT Optical Flow, U-Net Segmentation, ResNet, LSTM/RNN, Depth Estimation, PnP Pose Estimation, Object Detection

Dev Tools & Infrastructure: Docker, Git, Linux (Ubuntu), CI/CD, WPI Turing HPC (NVIDIA A30), NVIDIA RTX 4080 Simulation & Design: Gazebo, VizFlyt, Simulink, Fusion 360, ANSYS, Blender Experience

Research Engineer WPI Automata Lab (Prof. Kevin Leahy) Worcester, MA Jan 2025 – May 2026

Project: Real-Time Optical Communication System for Multi-Agent Robots Using Event-Based Vision

• Designed and built a high-frequency optical communication system using Prophesee EVK4 event cameras and LED signaling for robot-to-robot data transfer, achieving 95%+ decoding accuracy at 1200 bps during motion and 7x faster processing over prior EKF-based methods.

• Led hardware commissioning and sensor integration of event cameras with ROS2; developed software pipeline from raw event data through spatial filtering, blob tracking, and message decoding; validated system under varying motion speeds and lighting conditions.

• Implemented real-time blob tracking at 10,500 pixels/sec using UKF state estimation; developed spatial filter reducing 30K raw events per frame to 1K relevant events for low-latency decoding.

• Integrated AWS DeepRacer with ROS2 Nav2 for autonomous SLAM-based navigation; deployed GenX320 event cameras for real-time non-RF inter-robot perception and data exchange.

• Built simulation frameworks, benchmark plots, and system documentation to support validation and reproducibility; implemented bug fixes in production pipeline ahead of multi-robot field deployment. Research paper submitted to IEEE Robotics and Automation Letters (RA-L) (under review) Projects

SAFE NAVIGATION FOR ACKERMANN DRIVE ROBOTS — CBF-QP SAFETY FILTER ON AWS DEEPRACER

Python, C++, ROS2, OSQP, SICK LiDAR Mar – May 2026 https://github.com/pranaykatyal/safetybarrierondeepracer

• Commissioned an AWS DeepRacer with SICK TIM781S LiDAR and implemented a real-time CBF-QP safety filter on top of ROS2 Nav2, enforcing collision-free motion at 50Hz without replacing the nominal planner.

• Deployed two safety architectures: Point Cloud-Based CBF using k=7 nearest LiDAR returns as QP constraints, and Dynamic Parabolic CBF with online-adaptive boundaries for moving obstacle avoidance; achieved zero collisions across all test runs.

• Built and validated complete autonomy stack: LiDAR odometry (RF2O), SLAM, Nav2 path planner, CBF-QP safety node, and Dockerized deployment for reproducible lab and field testing. ROBOTIC ARM CONTROL — 4-DOF OPENMANIPULATORX

Python, C++, ROS2, Dynamixel SDK Nov – Dec 2024

• Led a 3-person team commissioning a full ROS2 kinematic and control stack for the 4-DOF OpenManipulatorX: implemented forward kinematics via DH-parameter homogeneous transforms, velocity kinematics as a ROS2 service computing the 6 4 Jacobian (Jv,i = Zi−1 (O4 − Oi−1), Jw,i = Zi−1), and inverse velocity using NumPy pseudo-inverse (np.linalg.pinv .

• Implemented PD controller (Kp=0.52, Kd=1.25) for Actuator 4 in Dynamixel effort (current) control mode via the SetCurrent interface at 1ms sampling rate; tuned gains through systematic hardware iteration — Kp=0.1 gave stability but high steady-state error, Kp=0.52/Kd=1.25 achieved 1.75% steady-state error with minimal overshoot across three reference positions (2000, 1600, 2100 encoder counts).

• Validated straight-line Cartesian motion in the +y direction by commanding constant end-effector twist ζ = [0, vy,0, 0, 0, 0]T through the Jacobian pseudo-inverse to joint velocities, then integrating qref = qref,old + q · tsampling at 0.1s intervals; logged position data to CSV confirming linear trajectory with minimal x/z drift. ADVANCED IK BENCHMARKING AND REAL-TIME ARM CONTROL — KINOVA GEN3 (6-DOF) MATLAB, Python, ROS2 Mar – May 2025

• Implemented and benchmarked five IK solvers (Newton-Raphson, Damped Least Squares, Gradient Descent, SVDDLS, IDLS, PSO) on a 6-DOF Kinova Gen3 arm in MATLAB; led SVDDLS implementation, translating the research algorithm into working code and enabling real-time VR-headset-driven arm control via Meta VR-derived poses.

• Modeled full arm dynamics using the Recursive Newton-Euler algorithm; generated quintic joint trajectories ensuring smooth, jerk-minimized motion between waypoints; analyzed forward kinematics via Product of Exponentials method.

• Conducted benchmark study comparing convergence speed, accuracy, and robustness across all solvers under varying initial conditions and singularity proximity; documented results in project report with quantitative comparison tables. QUADROTOR PATH PLANNING AND TRAJECTORY OPTIMIZATION Python, ROS2, VizFlyt Sep – Oct 2025

• Implemented RRT* path planner achieving 79% waypoint reduction (48 10 nodes) and 94.7% path efficiency; generated 7th-order polynomial minimum-snap trajectories satisfying kinodynamic continuity constraints through all waypoints.

• Designed cascaded PID control architecture (PX4-style: position velocity angular rate loops); achieved 5mm final positioning accuracy and 1.5cm average tracking error across 15.3s zero-collision autonomous missions.

• Deployed and validated on WPI Turing HPC Cluster (NVIDIA A30) in VizFlyt photorealistic simulation; minimum-snap optimizer produced dynamically feasible trajectories respecting quadrotor kinodynamic constraints at all waypoints. QUADROTOR CONTROLLER DESIGN AND HARDWARE VALIDATION — CRAZYFLIE 2.0 MATLAB, CrazyFlie 2.0 Nov – Dec 2024

• Designed and tuned PD and LQR controllers for the CrazyFlie 2.0 quadcopter in MATLAB/Simulink; modeled quadrotor dynamics and systematically swept gain parameters to characterize stability margins before hardware deployment.

• Validated both controllers on real CrazyFlie 2.0 hardware across 10-second path tracking benchmarks; LQR achieved 17.28% faster step response than PD, demonstrating improved settling time and reduced steady-state oscillation. MULTI-AGENT FORMATION CONTROL WITH SAFETY GUARANTEES (GCBFS) Python, ROS2, OSQP Aug – Dec 2025

• Designed and benchmarked synchronous (alpha=0.3) vs asynchronous (alpha=0.7) consensus protocols for a 5-drone pentagon formation tracking moving targets through obstacle fields (8m visual range, 6m communication range, 0.02s timestep).

• Implemented Graph-Based CBFs with distributed OSQP-based QP at O(n·k) complexity vs O(n2) centralized approaches; maintained 2.0m safety margins with zero CBF violations across all test scenarios.

• Demonstrated asynchronous protocol superiority: 50% lower control effort (0.024 vs 0.048 m/s2) and better formation tracking stability, while experiencing slightly higher formation drift (8.7m vs 7.5m); validated event camera-based non-RF communication as interference-free coordination channel. CHANCE-CONSTRAINED MOTION PLANNING WITH OMPL INTEGRATION C++, OMPL, Docker Mar – May 2025

• Integrated a Chance-Constrained RRT (CCRRT) algorithm with the Open Motion Planning Library (OMPL) in C++; introduced OMPL’s state space and planner interfaces into the CCRRT framework, led core algorithm development, and applied collision checking in configuration space for static obstacles.

• Achieved sub-1-second planning times under varying probabilistic safety constraints (δ {0.05, 0.10, 0.20}); evaluated how relaxing safety probability thresholds affects path length and planning time.

• Containerized full development environment in Docker for reproducible builds and deployment; contributed to project report and algorithm analysis.

AUTONOMOUS GAP NAVIGATION USING OPTICAL FLOW — DRONE PLATFORM RAFT, TS2P, ROS2, VizFlyt Nov – Dec 2025

• Designed a modular three-phase bidirectional navigation system for a simulated DJI Tello (100 100 200mm) traversing three rectangular windows (PnP-based detection) followed by one irregular gap (pure optical flow); full forward-and-return course without any depth sensor.

• Integrated GapFlyt’s TS2P (Temporally Stacked Spatial Parallax) on top of RAFT to amplify motion parallax across stacked frames, enabling reliable detection on featureless rear window faces where appearance-based methods fail entirely; fused dense RAFT predictions with temporally accumulated parallax cues for smoother, noise-resistant gap boundaries.

• Architected six-primitive skills-based FSM (SCAN, FIX YAW, ALIGN, VERIFY, APPROACH, RECENTER) in dedicated forward/return skill modules; computed camera intrinsics from simulated horizontal FOV using pinhole model

(fx = W/2 tan(FOVh/2)); reimplemented the provided controller from scratch after the original did not function correctly.

• Completed bidirectional 4-window course in 264.60 seconds simulation time with 0.729 seconds total inference; achieved 3rd place in class competition.

DEEP LEARNING WINDOW DETECTION AND DRONE NAVIGATION U-Net, PyTorch, Blender, ROS2, VizFlyt Sep – Oct 2025

• Trained U-Net semantic segmentation on 10,000 Blender-generated synthetic images with extensive augmentation pipeline

(Gaussian noise, blur, color jitter, elastic transform) for robust window frame detection under lighting and viewpoint variation.

• Deployed end-to-end pipeline on WPI Turing HPC (NVIDIA A30): segmentation PnP pose estimation depth-scaled visual servoing; achieved 1cm goal-reaching accuracy at 21ms inference (47Hz), completing a 3-window navigation course in 35 seconds.

• Designed synthetic data generation workflow in Blender with parametric scene variation (window geometry, materials, lighting, backgrounds) to maximize sim-to-real transfer for the perception stack. UNSCENTED KALMAN FILTER FOR IMU-BASED ATTITUDE ESTIMATION Python, MATLAB, 6-DOF IMU, Vicon Aug – Sep 2025

• Implemented full UKF for nonlinear 3D attitude estimation: 7D state (unit quaternion + angular velocity) with sigma points constructed in the reduced 6D space respecting the quaternion norm constraint; Cholesky decomposition with SVD fallback for numerical robustness; 12 sigma points propagated through quaternion integration process model.

• Solved the quaternion mean problem via iterative gradient descent: computed relative quaternions qrel,i = qi q, averaged rotation vectors, and applied adjustment quaternion iteratively until convergence — avoiding the discontinuity of naive quaternion averaging.

• Implemented sequential measurement updates: gyroscope model (with ZXY XYZ axis reordering for IMU mounting convention) then accelerometer model (gravity rotated via g = q gglobal q−1); Kalman gain K = Pxz(Pzz + R)−1 with separate Rgyro and Raccel matrices.

• Synchronized non-hardware-synchronized IMU (5,645 timestamps) and Vicon (5,561 timestamps) via SLERP; tuned Q = diag(3.4, 3.4, 3.4, 0.5, 0.5, 0.5), Rgyro = 15I, Raccel = diag(15, 15, 25) generalizing across 6 training and 4 blind test datasets; outperformed Complementary (α = 0.9995) and Madgwick (β = 0.01) filters. NEURAL NETWORK CONTROL ALLOCATOR FOR SHIP THRUSTERS PyTorch, NumPy, SciPy Aug – Sep 2025

• Reproduced the Skulstad et al. (2023) neural control allocation framework for a 3-thruster ship (T1: fixed bow, T2/T3: azimuth 180 ); implemented a two-layer LSTM encoder-decoder (hidden size 64) mapping 3D generalized force requests

(τx,τy,τψ) to 5D thruster commands (F1,F2,α2,F3,α3).

• Generated 1M synthetic training samples via random walk over thruster force/angle ranges (F1 [−10000, 10000] N, F2,3 [−5000, 5000] N, α2,3 [−180, 180], 70/15/15 train/val/test split); trained with AdamW + OneCycleLR (max lr=10−2, warmup 10%, 100 epochs, batch 256), dropout 0.2, gradient clipping 1.0.

• Composite loss L = k0L0 + k1L1 + k2L2 + k3L3 + k4L4 + k5L5 (physics MSE k0 = 10, autoencoder stability k1 = 1, magnitude k2 = 0.1, rate k3 = 0.01, power k4 = 0.001, forbidden sectors k5 = 0.1); achieved <6% error on sway (5.2%), yaw (4.9%), and mixed commands (5.4%); performance degraded on large surge (41.8%) due to T1’s fixed azimuth limiting surge authority. MONOCULAR DEPTH ESTIMATION WITH TRANSFER LEARNING

PyTorch, GPS/IMU fusion Nov – Dec 2025

• Reproduced M4Depth’s multi-scale temporal architecture (6-level feature pyramid with Domain-Invariant Normalization, temporal recurrence via feature warping, cost volume construction, and parallax-to-depth conversion) on WPI Turing HPC

(NVIDIA A30, 24GB); evaluated pretrained checkpoint on 550 MidAir synthetic validation samples, achieving 2.65m RMSE and 98.6% accuracy (δ < 1.25).

• Implemented custom TensorFlow dataloader for UseGeo real aerial dataset (7953 5279 images downsampled to 384 384, LiDAR-derived 32-bit TIFF depth maps, GPS/IMU pose CSVs); trained 146 epochs on 661 samples (Adam lr=0.0001, batch=3, sequence length=4), best checkpoint at epoch 80.

• Discovered critical pose format incompatibility: UseGeo provides absolute GPS/UTM coordinates (tx 498,340m) while M4Depth expects relative frame-to-frame motion (0–20m), causing parallax2depth to output depth estimates of 170–240 million meters; root-caused via value range checks and ground truth correlation analysis (r: −0.26 to 0.29).

• Implemented absolute-to-relative pose conversion (tcam = RT i−1trel, qrel = q i−1 qi) but fine-tuning still crashed at batch 47 with NaN loss; concluded pretrained weights encode strong scale assumptions incompatible with the converted pose distributions, a fundamental weight-level incompatibility beyond training instability. LSTM-BASED 3D DUBINS TRAJECTORY PREDICTION

PyTorch, NumPy Oct – Nov 2025

• Designed a rotation-invariant 3-layer LSTM (256 hidden units, seq2seq architecture) using body-frame representations to eliminate dependence on absolute heading; trained on 1,053,091 synthetic Dubins airplane trajectories with variable-length sequences and masked MSE loss to handle padding.

• Achieved 5.52m average prediction error ( 2%), 14.59m maximum error ( 6%), and 16.70m endpoint error across 100 diverse test cases with varying initial headings, climb angles, and path curvatures.

• Used 4D input (goal position + climb angle) to generate complete predicted trajectories; body-frame representation ensures the model generalizes across all orientations without requiring data augmentation over heading. VEHICLE CLASSIFICATION VIA RESNET18 TRANSFER LEARNING PyTorch, NVIDIA RTX 4080 Sep – Oct 2025

• Applied two-phase transfer learning with ImageNet-pretrained ResNet18 (11M parameters) for 10-class vehicle classification

(bus, sedan, SUV, truck, fire engine, etc.): Phase 1 froze backbone for 20 epochs (lr=10−3) to stabilize the new classification head to 93% accuracy; Phase 2 unfroze all layers for 30 epochs (lr=10−4, 10 reduction) to fine-tune pretrained features.

• Applied comprehensive data augmentation (RandomResizedCrop(224), RandomHorizontalFlip, ColorJitter, RandomRotation 10, ImageNet normalization) and Adam optimizer with weight decay 10−4 (L2 regularization) on only 1,400 training images; demonstrated that proper augmentation strategy eliminates overfitting without dropout.

• Achieved 96.0% validation accuracy (192/200 correct, loss 0.13) on NVIDIA RTX 4080 in 22 minutes; validation (96.0%) slightly exceeded training (95.4%), confirming no overfitting; identified that extending Phase 1 from 10 20 epochs was critical, yielding 8.5pp improvement; analyzed failure cases (green fire engine misclassified as heavy truck due to color distribution in training set).

ROBOTIC DOG STRUCTURAL DESIGN AND LOAD ANALYSIS (SPOT) Fusion 360, ANSYS Jan – Mar 2024

• Designed a quadruped robot (SPOT-inspired) in Fusion 360, modelling joint geometry, leg linkages, and frame assembly; performed ANSYS finite element load analysis under static loads up to 1000g, confirming structural integrity and informing material selection and joint reinforcement decisions.

• Iterated on joint and frame geometry based on FEA stress distributions to ensure reliable performance across gait cycles; documented design rationale and simulation methodology in project report. Leadership

Secretary IEEE-RAS WPI Chapter Worcester, MA

Jan 2025 – Present

• Coordinated industry speaker events, technical workshops, and robotics demonstrations for 50+ student members.

• Managed communications between faculty, industry partners, and student members; grew chapter participation by 30%.

Contact this candidate