Back Home

Portfolio for UMA

See also: main projects page.

“We sail with in a vast sphere, ever drifting in uncertainty,” wrote Pascal 400 years ago. Nothing excites me more. He was writing about humankind, but he could’ve easily been describing robotics. You’re looking for roboticists with “a strong drive to get humanoid robots working in the messy real world.” Nothing thrills me more about robotics: Building intelligent machines that can adapt and thrive in uncertain environments.

My favorite memories as an engineer have been out in the field: Analyzing point clouds under an oak tree, testing autonomous cars in the early morning, running spare batteries from the lab to the streets. In these real-world deployments, I quickly discovered edge cases that I never would’ve dreamed of, often patching them in the field. For the reforestation robot that I helped develop at Carnegie Mellon, I needed to build a planner than accounts for aggressive deer, for example.

As an engineer at the Robotics and AI Institute, I’ve built up and maintained a fleet of 24+ robotic arms to support 5+ research groups. I’ve designed and built a state-of-the-art force-mediated teleop device so that we can train the very best force-aware policies. I’ve built out and deployed an industrial-grade decentralized diagnostics system, including a full-stack web app, that has cut our debugging time from hours to seconds by inspecting hundreds of datapoints in real time.

I want to work at UMA because I believe in the core promise of autonomy: Enhancing lives for everyone. But that can only happen when we build robots that don’t just survive in uncertain environments but thrive in them.

Manipulation engineering at RAI Institute (2025 — Present)

With the arms at RAI

Since June 2025, I’ve served as the operations engineer of a manipulation lab with 24+ Franka robot arms supporting five research groups. While much of my current role is more infrastructure-focused, it keeps me sharp on the integration and debugging skills that matter for Stretch. I’ve also developed novel hardware and software tools that researchers have come to depend on.

Force-aware teleoperation device

I developed a novel force-aware leader arm teleoperation device with gravity compensation, force feedback, and programmatic friction compensation. It’s a 7 DOF leader arm that feels effortless to move, running through a high-frequency controller with optimized inverse dynamics calculations. We’re in the process of open-sourcing the hardware and software. It’s now used by nearly every manipulation research group at RAI, and is currently being extended to new embodiments beyond Frankas.

Central diagnostics stack

And I built the lab’s core diagnostics system: decentralized agents across a dozen machines, a full-stack web dashboard with interactive 3D visuals, and centralized logging — the kind of tooling that lets you trace a vague hardware symptom to a specific subsystem quickly.

Cleaning up our operations

I spearheaded a massive standardization initiative that unified several distinct research testbeds, each with their own manipulators, sensors, compute, and networking, into a powerful, clearly defined standard. This meant gathering feedback from stakeholders across five research groups, who all share the same testbeds.

Reforestation robot at CMU (2023 — 2025)

(CMU news release)

From 2023–2025, I served as project manager and software lead for an award-winning reforestation robot in the Kantor Lab at CMU’s Robotics Institute. I architected and implemented the complete robotics software stack in ROS2, spanning perception, behavior planning, navigation, manipulation, and ecological modeling.

The software diagram for Jonny, the second reforestation robot that I helped build at CMU. It includes the Behavior Manager (FSM), Forest Planner and Waypoint Manager (high-level waypoint generation), Trajectory Planner (using costmaps), and local control using Pure Pursuit.
Canopy's robot control section includes a global map and a local 3D view of the robot, including the navigation system's costmap and trajectory. The green checkmark shows that all diagnostic checks have passed. The dot below the check flashes to indicate that the telemetry is live. Depicted is the first reforestation robot that I helped develop, Steward.

The software diagram for Jonny, the second reforestation robot that I helped build at CMU. It includes the Behavior Manager (FSM), Forest Planner and Waypoint Manager (high-level waypoint generation), Trajectory Planner (using costmaps), and local control using Pure Pursuit.

Behavior trees & task planning

The robot’s autonomy was built on behavior trees that coordinated a complex task sequence: navigate to a planting site, align with the GPS waypoint, lower the auger, drill, deposit a seedling, and move on — while handling faults like soil obstructions, GPS drift, and mechanical jams. The behavior system managed transitions between autonomous planting, assisted teleoperation, and manual override.

Manipulation & grasping

The planting mechanism required precise coordination between a robotic arm and a custom auger. I developed the control sequencing for the manipulation pipeline, including force-based detection of soil contact and obstruction handling.

Our second reforestation robot, Johnny, drills a hole and plants a seed using an xArm with a custom, low-cost dual end effector that I designed. The planting sequence is guided by an FSM. The drilled hole is aligned with the grasper using an eye-in-hand stereo camera
The custom, low-cost end effector that I designed and manufactured. The brush, used to gently compact the soil around the tree seedling, was sawed off of a lab broom. The 'hand' is a simple parallel gripper powered by a Dynamixel. We had no budget!

Our second reforestation robot, Johnny, drills a hole and plants a seed using an xArm with a custom, low-cost dual end effector that I designed. The planting sequence is guided by an FSM. The drilled hole is aligned with the grasper using an eye-in-hand stereo camera

Navigation & localization

I implemented GPS-guided waypoint navigation with local obstacle avoidance. The system operated in unstructured outdoor terrain — grass, slopes, tree roots — where clean odometry is a luxury.

Assisted teleoperation

I implemented a hybrid control scheme for teleop in the field, and validated the scheme with a user study. The method blends direct joystick inputs, which I assume to be noisy, with fused perception data structured as a costmap, to produce adjusted teleop commands guaranteed to be safe.

Early tests with my custom-built full-stack control dashboard (left) and my custom-built, Unity-based simulator (right).
Diagram of the assisted teleop system, including noisy user inputs and a costmap C.
Qualititative results from the user study show a clear preference for the assisted teleop system over direct control.

Early tests with my custom-built full-stack control dashboard (left) and my custom-built, Unity-based simulator (right).

Simulation & HIL testing

I developed Ecosim, a custom photorealistic farm robotics simulator in Unity that served as a drop-in replacement for hardware telemetry and actuation. This enabled hardware-in-the-loop testing and validation without needing the physical robot, which was critical for rapid iteration on behavior logic.

Canopy is a full-stack forest planner. Users select the trees they'd like to plant-- guided by climate-appropriate recommendations-- and simply paint them onto the map. The precise species mix and distribution is determined by an RL policy that I developed in the backend and trained on a forest growth model, with near-instant inference times. Planning a forest is as simple as finger painting.
Canopy's design was informed and validated by real foresters. Here, a forester from the Pennsylvania Bureau of Forestry-- who had zero prior robotics experience-- painted and uploaded a reforestation plan to the robot in minutes as part of a public deployment.
To train and test the software, I built a digital twin of our deployment area using Blender and Unity.

Canopy is a full-stack forest planner. Users select the trees they'd like to plant-- guided by climate-appropriate recommendations-- and simply paint them onto the map. The precise species mix and distribution is determined by an RL policy that I developed in the backend and trained on a forest growth model, with near-instant inference times. Planning a forest is as simple as finger painting.

Results

The robot won the Excellence in Regenerative Agriculture award at the 2025 Farm Robotics Competition, the largest competition of its kind in the world, and was invited to demo at FIRA 2025. We deployed the robot for a public planting project in Schenley Park in partnership with state and local officials.

Autonomous driving platform at UT Dallas (2019 — 2023)

From 2019–2023, I founded and led an autonomous driving research project at UT Dallas that culminated in a publicly deployed self-driving car. I served as the project’s lead and software architect, and this is where I cut my teeth as a robotics generalist.

The system diagram for Navigator, our open-source AV stack

Behavior planning & task management

I designed and implemented the vehicle’s behavior management system in C++, using finite state machines to coordinate driving modes, safety states, and fault recovery. The system handled transitions between autonomous driving, manual override, emergency stops, and degraded-capability modes— all with hard real-time constraints because a 2-ton vehicle doesn’t wait for your code to catch up.

Navigation & path planning

I implemented the full navigation stack: an HD (semantic) mapping system for route-level planning, and local planners using A*, RRT*, and lattice-based approaches for real-time trajectory generation. Route candidates and trajectories were visualized in real time through a custom dashboard (three.js) mounted on the vehicle’s touchscreen.

Perception, localization, and mapping

Real-time SLAM pipeline

I implemented and tuned SLAM algorithms using GTSAM, NDT, and KISS-ICP. I built camera and LiDAR sensor fusion pipelines, including octree-based spatial indexing and planar segmentation with RANSAC for ground filtering. Point cloud and occupancy grid data fed directly into the planning stack. I implemented an Extended Kalman Filter that fused RTK data with wheel odometry and IMU measurements to provide centimeter-level accuracy, even when GPS signal was obstructed.

Low-level systems & hardware integration

Real-time SLAM pipeline

I wrote low-level C++ interfaces to safety-critical system, including to CAN buses for direct control of the vehicle’s pedals and steering. I built high-performance sensor filters and developed custom binary protocols for streaming LiDAR packets over WiFi in real time.

Debugging complex cyberphysical systems

Deploying a self-driving car on public roads means debugging at every layer of the stack. I’ve profiled rendering pipelines in a custom-built simulator, traced faulty terminating resistors in CAN networks with a multimeter and oscilloscope, and diagnosed timing issues across distributed ROS nodes. When something goes wrong at 30 mph, you learn to be systematic.

Technical depth

C++: Safety-critical CAN bus interfaces, high-performance sensor filters, behavior state machines, firmware for sensors and actuators. I’m comfortable working day-to-day in a C++ codebase.

Robotics generalist: Behavior trees and FSMs, grasping and manipulation, navigation and path planning (A*, RRT*, lattice planners), SLAM and localization (GTSAM, NDT, KISS-ICP), perception and sensor fusion (camera, LiDAR, octrees, RANSAC), hardware debugging with oscilloscopes and multimeters.

Integration: I’ve built systems where perception, planning, control, and hardware all have to work together in real time — on a self-driving car, on a field robot, and across a fleet of 24 research arms. I enjoy the puzzle of making it all fit.

ROS/ROS2: Used extensively across both the reforestation project and lab operations at RAI. Docker, Bazel, and CI/CD for deployment.

Python: Web servers, SLAM implementations, CNNs, automation tooling. I reach for Python when C++ isn’t needed and when rapid iteration matters.

The software diagram for Jonny, the second reforestation robot that I helped build at CMU.
The software diagram for Navigator, the open-source autonomous driving stack that I managed.

The software diagram for Jonny, the second reforestation robot that I helped build at CMU.