robotics software engineer
Drivers, control loops, teleoperation, safety systems and the pipelines that turn recorded episodes into trained policies. Two robot stacks built end to end; one of them in the field.
Scrollthe stack
Every box is code I wrote or own on a robot arm that is in the field today.
teleoperation
At Open Droids I built a VR teleoperation pipeline that streams a robot's stereo depth camera to a Meta Quest 3: 100–200 ms latency on the local network, 300–400 ms with the robot in San Francisco and me in Ottawa. At Sielo I integrated four input devices for the same arm, Quest 3S controllers, a SpaceMouse, the wheelchair's own joystick and a browser joystick, behind a joystick arming interlock, plus teach-by-demonstration record and replay and a zero-torque hand-guiding mode.
control and safety
The same arm over a year. In between: a C++ ros2_control hardware interface and CAN driver for seven actuators from two vendors; a 250–400 Hz loop with per-joint PD gains, feedforward, filtered feedback and gravity compensation; control and teleop processes on SCHED_FIFO with CPU pinning on a PREEMPT_RT kernel, which removed a power-on jerk; and a layered safety system: hardware E-stop, ESP32 firmware with a 20 Hz heartbeat and 150 ms failsafe, a velocity guard at the CAN boundary and a torque-residual collision detector.
perception and assisted control
Click-to-go back-projects a clicked pixel through depth into the robot's base frame. A YOLOv8 pipeline on an OAK-D camera turns detections into 3D grasp targets through a calibrated camera-to-base extrinsic.
robot learning data
SieloStudio, which I wrote alone, records teleoperated episodes to MCAP, converts them to LeRobot datasets, trains ACT, Diffusion Policy, SmolVLA and π0, and pins every model to a data-contract version so a deployed policy sees the same inputs it was trained on. π0, π0.5 and Diffusion Policy have run on the real arm through the Rosetta LeRobot-to-ROS 2 bridge.
An off-process recorder reads raw messages without decoding them and cut recording CPU from 0.76 to 0.10 cores. The machine-tending rig above runs the same kind of stack on a second arm: perception from the wrist camera, and the door, the part and the bins handled without an operator.
arlo, sielo robotics, 2025–
6-axis assistive arm on NVIDIA Jetson. About 90% of the commits are mine: driver, control, safety, teleop, perception, data pipeline, fleet tooling with health-checked over-the-air updates and automatic rollback, and the architecture guide, about 1,900 lines.
r1d1, open droids, 2024–25
Founding engineer. Off-the-shelf components from several vendors made to work as one robot: interfaces for their RS232 and RS485 protocols on a Jetson Orin NX, and the ROS 2 layer for control, planning, perception and navigation. At CES 2025 the robot detected an object, estimated its 3D position from depth, planned a grasp with MoveIt and handed it to a person.
projects
Kinematics derived from scratch, custom ROS 2 controller, tested in Ignition Gazebo. 30,000+ views on LinkedIn, reposted by Open Robotics.
Isaac Sim environment connected to ROS 2 for state, perception, planning and control.
ACT, Diffusion Policy and π0.5 on a LeRobot arm.
Eight-wheel agricultural UGV: ArduPilot, Jetson Nano, RTK GPS. INR 250,000 in government funding; Robofest finalist.
about
Robotics software engineer in Ottawa. M.Eng. Electrical and Computer Engineering, University of Ottawa, 2024–2026. B.E. Mechatronics, Gujarat Technological University, 2019–2023. Python and C++ on Linux; ROS 2; NVIDIA Jetson; Isaac Sim and Gazebo.
contact
Email is fastest, usually answered the same day. patelyugmil@gmail.com · LinkedIn · Ottawa, ON.
