Featured Projects
Discrete Event Simulation (DES) for Robotics
Built a high-speed Discrete Event Simulation engine for multi-robot systems, delivering results in seconds instead of hours and enabling rapid design space exploration, What-If scenario analysis, and Monte Carlo-driven reliability estimation.
Challenges
- Traditional time-based robotics simulators took prohibitively long to run, making iterative system design and layout exploration impractical
- Evaluating different design configurations (fleet size, facility layout, task allocation) required running full physics-based simulations each time
- Computing reliability metrics like p95 or p99 performance required thousands of simulation runs, infeasible with existing tools
- No fast way to perform What-If scenario analysis for varying operational configurations across logistics, warehousing, and industrial automation domains
Solutions
- Developed a Discrete Event Simulation (DES) engine that models queues, resources, and decisions instead of continuous physical time
- Built automated layout parser converting facility designs to graph representations for simulation
- Enabled Monte Carlo simulations for design space exploration and p95/p99 reliability calculations
- Implemented visualization engine for simulation results and robot path analysis
- Designed a fully configurable system supporting robot parameters, layout parameters, and operational constraints
Outcomes
- Over 3,600x speedup — simulations complete in seconds vs hours with traditional time-based tools
- Less than 3% error compared to time-based simulation benchmarks, validated across multiple configurations
- Enabled rapid evaluation of fleet sizing and layout alternatives that were previously infeasible to explore
- Unlocked Monte Carlo-driven design space exploration for robust system design
- Paved the way for Multi-Agent Path Finding (MAPF), RL-based layout optimization, and online forward simulation
Physics-Informed Neural Networks for Satellite Simulation
Accelerated satellite orbital simulation using Physics-Informed Neural Networks (PINN) and Physics-Informed Recurrent Neural Networks (PI-RNN), eliminating dependency on expensive synthetic data generation while ensuring physical consistency.
Challenges
- Generating large volumes of synthetic orbital data using standard ODE solvers (RK4) was computationally expensive and time-intensive
- Models trained only on synthetic data failed to adapt to real satellite telemetry due to unmodeled effects (solar radiation pressure, atmospheric drag, magnetic perturbations)
- This created a modeling gap where trained models performed well on simulation but drifted when applied to real-world conditions
Solutions
- Implemented Physics-Informed Neural Network (PINN) to embed physical laws directly into the learning process
- Trained PINN simultaneously on real telemetry and physics-based loss, eliminating need for large synthetic datasets
- Used dummy collocation points to enforce physics consistency without costly data generation
- Developed Physics-Informed Recurrent Neural Network (PI-RNN) for orbital propagation, capturing temporal dynamics and handling maneuvers as control inputs
Outcomes
- Synthetic data generation became feasible on standard laptop GPU (NVIDIA RTX) - no need for high-end compute clusters
- Successfully merged real-world telemetry with physical constraints, leading to more stable and physically consistent synthetic data
- Maintained physical validity while reducing computational costs
- Enabled faster iteration cycles for model development and testing
Post-Maneuver Orbit Prediction & Tracking
Developed a probabilistic framework for predicting satellite orbits after unobserved maneuvers by progressively constraining an infinite maneuver space into a feasible prior distribution, then converging to precise orbit determination using Bayesian filtering and multi-hypothesis tracking.
Challenges
- After unobserved maneuvers, the satellite's new orbit is unknown — modeling all possible maneuvers produces an infinite-dimensional search space in RTN (Radial-Tangential-Normal) coordinates
- Brute-force exploration of the unconstrained maneuver space was computationally infeasible for real-time tracking and sensor tasking
- Ambiguous sensor measurements (radar, optical) with multiple unidentified objects required robust data association to link detections to the correct satellite state
- Needed to integrate physics constraints (two-body dynamics, J2 perturbations, drag, solar radiation pressure) with operational priors (fuel budgets, thruster hardware limits, mission intent)
Solutions
- Designed a 6-step progressive constraint pipeline: hardware magnitude bounds, fuel budget via Tsiolkovsky equation with LogNormal priors, anisotropic directional priors favoring tangential burns, timing/efficiency priors at optimal orbital phases, and mission-context weighting
- Implemented Monte Carlo profile generation using Dirichlet-distributed budget splitting across maneuver segments with configurable RTN direction sampling
- Built an orbital propagator integrating impulsive and finite-duration burns with RTN-to-ECI frame transformations for each maneuver segment
- Applied Unscented Kalman Filter (UKF) for Bayesian state estimation, collapsing the uncertainty cloud as measurements arrive
- Developed Joint Probabilistic Data Association (JPDA) and Multiple Hypothesis Tracking (MHT) running parallel UKFs to handle ambiguous multi-target measurement scenarios
Outcomes
- 99.99% reduction in search volume — from over 10 million km³ (unconstrained) to under 1,500 km³ through progressive constraint application
- Each constraint layer provided measurable reduction: fuel budget prior (~100% from baseline), individual ΔV prior (48.4% further), directional prior (25.8% further)
- Probabilistic framework enables rapid sensor tasking by focusing search on high-likelihood orbital regions rather than brute-force sky surveys
- Multi-hypothesis tracking (JPDA/MHT) provides robust orbit determination even with sparse, cluttered, or ambiguous measurements
- Implemented using Poliastro, Astropy, NumPy, and SciPy for production-grade orbital mechanics computations
Control Systems & State Estimation
Advanced control algorithms and state estimation techniques for autonomous systems, including Kalman filtering, model predictive control, and adaptive control strategies for dynamic environments.
Challenges
- Handling uncertainty and disturbances in real-world autonomous operations
- Achieving optimal performance while guaranteeing safety constraints
- Adapting control strategies to varying operational scenarios and environmental conditions
Solutions
- Implemented Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) for nonlinear state estimation
- Developed model predictive control (MPC) with real-time constraint handling
- Created adaptive control algorithms that learn and adjust to changing system dynamics
- Built hierarchical decision-making architecture for mission planning and execution
Outcomes
- Robust state estimation even with noisy, uncertain measurements
- Smooth, precise autonomous operation in challenging conditions
- 100% safety compliance with zero constraint violations
- Successful deployment in GPS-denied and dynamic environments