Farzad
Ahmadinejad

Autonomous last-mile delivery robots

Lomby · Tech Lead · 2024 to now

I lead a team of six engineers and own the software architecture for the platform. We replaced our off-the-shelf navigation with an in-house ROS 2 stack built for the specific problem of narrow Japanese sidewalks shared with pedestrians: point cloud processing, localization, planning, and control.

I also built the on-robot telemetry module that records what happened on every delivery and uploads it automatically, which is how we find out what the fleet actually does rather than what we think it does.

30%Faster average delivery
50%Less operator intervention
6Engineers led

• ROS 2
• C++
• Python
• 3D LiDAR
• Sensor fusion
• Fleet telemetry

Omnidirectional AGV, and a robot that follows you

Blue Innovation, MehrTronix · 2016 to 2020

Low-level firmware, ROS controller, 2D LiDAR SLAM, and navigation for a Mecanum-wheeled AGV that worked in coordination with a drone. Earlier, at MehrTronix, I led a human-following AGV built with Kyutech University, using transfer-learned detection on MobileNet-SSD and OpenVINO.

• ROS
• Firmware
• 2D SLAM
• OpenVINO
• Mecanum kinematics

Close-proximity power line tracking

Japan Infra Waymark (NTT West) · 2020 to 2024

One of four custom inspection drones I built in this role. This one flies alongside live overhead cables and holds position within ten centimetres of them. Target detection through PCL segmentation, SLAM, and autonomous flight control, fusing RTK GPS with 3D LiDAR on ROS 2. Ten centimetres is close enough that every part of the estimation chain has to be right at once.

10 cmStandoff from a live cable
1 of 4Custom inspection drones built

• ROS 2
• PCL
• RTK GPS
• 3D LiDAR
• SLAM

Flight without GPS

Blue Innovation · 2019 to 2020

3D mapping, path planning, and navigation for drones indoors, built on VIO and 2D LiDAR with an Octree map, JPS planning, and MPC for tracking. Shown at CES 2020. My master's research went the other direction, into the ground effect that makes quadcopters unstable near surfaces, which is a problem you meet immediately once you fly indoors.

• VIO
• Octree
• JPS
• MPC
• C++

Omnidirectional USVs for bridge inspection

Japan Infra Waymark (NTT West) · 2020 to 2024

Four generations of unmanned surface vessel, from the electronics subsystems and controller firmware up to an autonomous navigation system fusing a VIO camera, 3D LiDAR, and GNSS. Under a bridge deck you lose GNSS exactly when you need it most, and the water surface gives a LiDAR almost nothing to hold onto, so the hard part is knowing where you are at all.

Eight boats went into service. They have been inspecting bridges for over five years.

8Vessels in service
5+ yrsContinuous field use
4Hardware generations

• ROS 2
• VIO
• 3D LiDAR
• GNSS
• Firmware
• PCB design

Humanoids, manipulators, and a parcel sorter

Amirkabir University, MehrTronix · 2011 to 2019

Kid-size humanoid soccer robots, both custom-built and NAO, where I worked on bring-up and gait, including an evolutionary method for tuning an offline walk engine that we published. Later, a 5-DoF manipulator for RoboCup@Work with inverse kinematics, CNN-based 3D object detection, and grasping, and custom behaviours on Pepper.

Also a computer vision system that reads shipping labels off parcels moving at a metre per second on a conveyor, built for Digikala and published at ICSPIS 2018.

• Inverse kinematics
• NAO
• Pepper SDK
• CNN detection
• Grasping