Robotics studies developed as engineering investigations rather than software demonstrations: models and assumptions are made explicit, competing methods are evaluated under controlled conditions, and conclusions are tied to quantitative evidence, implementation behavior, and physical constraints.
Engineering investigations spanning robot dynamics, model-based control, reinforcement learning, ROS 2 runtime validation, motion planning, perception, and sensor fusion. Each project emphasizes physical assumptions, quantitative evaluation, and reproducible implementation.
Examines bounded residual torque policies under payload-model mismatch using PPO, SAC, and TD3. PhysX experiments test heavier and lighter payloads, reverse-mismatch learning, and 4 kg cross-evaluation to identify when learned compensation improves—or degrades—an established nominal controller.
Develops a physically constrained three-second Cartesian transportation benchmark and compares computed-torque PD with constrained MPC. ROS 2–Isaac Sim validation evaluates G025 tracking, gravity compensation, startup stabilization, and execution timing before introducing payload uncertainty.
Reconstructs and checks the six-joint UR5e model against Isaac Sim, including coordinate frames, forward and inverse kinematics, Jacobian conditioning, rigid-body dynamics, payload inertia, and Tool0 center-of-mass consistency as the foundation for physically interpretable control.
A Cartesian trajectory-tracking study comparing exact-model and mismatched computed-torque control, Pure SAC, and CTC + Residual SAC. A controlled 30% link-2 mass error is traced from model mismatch to analytical torque deficit, then used to test whether a bounded learned residual can recover the missing dynamics without discarding the structure of the classical controller.
Tests whether analytically validated controllers retain their behavior after crossing the simulation and software boundary. Independent Isaac Sim validation and ROS 2/C++ execution are used to examine model-fidelity gaps, control-loop timing, computation latency, actuator interaction, MPC deadline and completion failures, and the distinction between a mathematically valid design and a practically executable controller.
Compares PID, computed-torque, LQR, and MPC as different control architectures rather than as a single ranking exercise. Nominal tracking is coupled with model mismatch, disturbance rejection, gravity-feedforward structure, constraint handling, horizon and tolerance sensitivity, computational cost, and Python–C++ cross-validation to identify where each method's assumptions become operationally significant.
Quantifies the nonlinear structure that governs achievable motion and control authority. The study examines Jacobian conditioning and singularity behavior, pseudoinverse versus damped least-squares inversion, configuration-dependent inertia and gravity loading, inverse dynamics, torque demand, local linearization validity, and trajectory-level numerical consistency across Python and C++ implementations.
Establishes a defensible plant model before controller design by reconciling nominal DH geometry, URDF frames, inertial parameters, fixed-base articulation, and Isaac Sim configuration. Forward kinematics, Jacobians, mass matrices, and gravity generalized forces are independently cross-checked to expose frame, parameter, and simulator inconsistencies rather than absorbing them into later tuning.
A controlled benchmark that distinguishes search-algorithm behavior from robot-specific feasibility. Dijkstra, A*, RRT, and RRT* are evaluated across discrete and metric environments, then connected to footprint inflation, continuous clearance, stochastic convergence, trajectory refinement, proximity-aware velocity regulation, and mission-level execution.
A geometry-driven fusion pipeline built around calibrated frame transforms, 3D-to-2D projection, object-level depth association, YOLO-World, and FastSAM. Controlled perturbation studies isolate sensitivity to extrinsic yaw, viewpoint change, and range bias so fusion robustness is evaluated against identifiable geometric error sources rather than visual output alone.
An integrated detection, persistent-tracking, and instance-segmentation pipeline evaluated as a real-time system rather than a model demo. Frame-level latency, throughput, processing-time distribution, object-count variation, and scene-load effects are measured to separate neural inference cost from end-to-end perception behavior.
A nonlinear state-estimation study fusing wheel odometry and IMU measurements in ROS 2/Gazebo. Process and measurement uncertainty are treated explicitly, and the fused estimate is evaluated against ground truth to quantify trajectory error, RMSE reduction, and the remaining drift expected when no absolute-position observation is available.