My first arm to use NEMA 17 steppers and a belt-driven elbow, with 3:1 planetary reductions (not my own design yet).
Portfolio · 2026
Mechatronics & software
Logan Vimalaraj
I build robots from the gearbox up: cycloidal drives, firmware, and the software that lets them think.
Now building: DUM-E(About)
I’m a student at the Davidson Academy who likes to own every layer of a machine: the CAD, the reductions, the firmware and the ROS stack on top. Each project teaches me something the next one needs.
01 The manipulator
Robot Arm v7
The seventh version of my arm, and the one DUM-E carries. I designed every part from scratch in Onshape.
- Degrees of freedom
- 6+ gripper
- Reach
- ~500mm
- Payload
- ~800g
- Cycloidal reduction
- 15: 1
Specification
- Wrist
- RPR: a roll joint, then a differential for pitch and roll
- Motors
- NEMA 23 on the shoulder; NEMA 17 on the base, elbow and wrist
- Feedback
- Closed loop on the base, shoulder and elbow
- Reduction
- Custom low-profile cycloidal drives on the joints
- Structure
- 2020 aluminium extrusion segments with 3D-printed parts
- Gripper
- DS3218MG servo
- Control
- ESP32 bridge to a ROS 2 master, such as a Raspberry Pi
The 15:1 cycloidal gearbox is open source. Download it on MakerWorld ↗
02 Flagship project
Hardware complete · Software in progress
DUM-E
A general-purpose home robot. Give it a long-horizon command like “clean this room” and it plans the steps, drives itself where it needs to be, and does the work with its arm.
Drivebase + vertical lift + 6-DOF arm
All three now built and working together.
Project page →
Arm
Lift
Drivebase
Raspberry Pi 4 inside
-
01
A 6-DOF arm
The arm is my Robot Arm v7: six degrees of freedom plus a gripper, custom cycloidal drives in the joints, and closed-loop control on the base, shoulder and elbow.
Robot Arm v7 in detail -
02
A vertical lift
The whole arm rides a stepper-driven carriage on an aluminium mast, so the same arm can work near the floor or reach up to desk height.
-
03
A drivebase
An aluminium-skinned mobile base carries the tower and its electronics around the house. The LLM planner and SLAM stack steer it directly.
-
04
A brain
A Raspberry Pi 4 inside the white base is the onboard computer and ROS 2 master, with an ESP32 bridging it to the motors. On top, an LLM + pointer + SLAM planner breaks a command into steps and a VLA model handles precise arm control.
Software stack
From one sentence to motor steps
- Input “Clean this room” Long-horizon text command
- Plan LLM + pointer Breaks the task into smaller steps
- Navigate SLAM Drives the base to each step
- Manipulate VLA policy Precise control of the arm
- Actuate ROS 2 → ESP32 Joint commands to the steppers
03 Lineage
Each arm taught the next one.
The first arm with actuators I designed entirely myself. Every joint runs my 9:1 cycloidal drive.
Aluminium extrusion segments made a huge 800 mm reach possible. A 64:1 three-stage planetary drives the shoulder and 9:1 cycloidals drive the elbow and base.
Smaller than v5 but with far more torque headroom. Stackable 5:1 planetary gearboxes, doubled up to 25:1 at the shoulder and base.
5 DOF plus gripper. A fiberglass-reinforced GT2 belt drives the elbow, and the wrist uses differential helical bevel gears. Read more
6 DOF plus gripper with custom cycloidal drives and closed-loop joints. It’s the arm that rides on DUM-E. Back to v7
04 Simulation
A digital twin, driven by Claude.
DUM-E in MuJoCo, speaking the same ROS 2 topics as the real robot. Claude Opus 5.5 runs it end to end through an MCP server.
Run 01Test room
A room full of chores.
Stack and sort cubes, bring cans down from a shelf (the top one needs the lift), open a drawer and lift out a block, press a button, then park on the charger. Stacking and sorting share the same cubes, so 7 of 8 is the most the board can show at once. Claude did both.
- Tasks
- 7/ 8
- Real time
- 41min
- Shown at
- 35×
Run 02One-story house
Five rooms, no coordinates.
A navigation test. Claude gets directions in words, never positions or ground truth: only drifting odometry, LiDAR, ToF sensors and the cameras. It toured every room, carried a soda can from the kitchen to the office desk and a towel to the laundry basket, switched on a bedside lamp, and docked, with 12 bumps along the way.
- Tasks
- 5/ 5
- Real time
- 37min
- Shown at
- 25×
How it works
- Physics
- MuJoCo, with the drive ESP32, arm bridge, lift, IMU, ToF, cameras and LiDAR emulated
- Interface
- The real robot’s ROS 2 topics, so DUM-E’s own control stack runs on it unchanged
- Agent
- Claude Opus 5.5 over MCP: observe, drive, move the arm, set the lift and gripper
- Judging
- The sim scores each task from the physical state and counts every bump
The simulator, task boards and MCP server are on GitHub. loganoe/simulator ↗
05 More work
Other things I’ve made
-
A personalized learning platform built with Justin Baratta. It uses an LLM to turn a text prompt and a difficulty setting into a full course, with lessons, examples, sources and practice problems.
Winner of the 2025 Congressional App Challenge for Nevada’s 2nd District.
-
A compact, 3D-printed 9:1 cycloidal reduction for a NEMA 17 stepper, designed in Onshape for my arm actuators. Two cycloidal disks run out of phase to cancel vibration.
![3D-printed 9:1 cycloidal drive gearbox for a NEMA 17 stepper motor]()
-
Built with Siddhartha Hiremath (sdrt.dev). A Cartesian gantry driven by NEMA 17 motors moves a magnet on a linear actuator under the board, so it can play a physical game of chess against you.
-
My best pen plotter so far. It moved from 28BYJ-48 motors to two NEMA 17 steppers on a Cartesian XY system, with a fully 3D-printed frame.
06 About
Hardware, firmware & everything between
I’m Logan Vimalaraj, a student mechatronics and software developer based in Reno, Nevada, studying at the Davidson Academy. I connect robotics hardware, low-level firmware, ROS and full-stack software into complete machines.
- Recognition
- Winner, 2025 Congressional App Challenge (NV-02) for Insight. Announcement ↗ House.gov ↗
- Tools
- Onshape, 3D printing, stepper and servo control, ESP32, ROS 2, SmolVLA, LLMs
- Elsewhere
- GitHub ↗