Jewel Nath
← Home

Robot Niko

A life-size, 3D-printed humanoid assistant, built in 2022 with my team at Quanta Robotics. It wakes to "Hey Niko", answers questions with a jaw that moves as it speaks, turns its head to follow a face, and greets the people it knows with a handshake.

Built2022My part: every part of it: the programming, the 3D printing, the assembly, the servo tuning and the arm motions

What it shows

Motion
27 servos on two PCA9685 boards, driven from one Raspberry Pi as one set of channels, with a library of gestures
Voice
An on-device wake word, an online assistant, and its own question bank, with a jaw that moves as it speaks
Vision
OpenCV face detection that steers the head, and face recognition that greets people by name
The build
InMoov's open design, 3D-printed, assembled and wired by the team, from a 3D model in August to a finished robot in November

In video · December 2022

Niko working

Recorded at the FabLab in the days around its launch, with sound.

6 Dec 2022 · 1 min 44 s

Two people come up to Niko one at a time. It greets each one and shakes their hand.

6 Dec 2022 · 53 s

Its head turns to follow the face of a person walking left and right in front of it.

On YouTube

7 Dec 2022 · 41 s

Niko introduces itself.

On YouTube

7 Dec 2022 · 3 min

Two of the team ask Niko questions, and it answers them.

7 Dec 2022 · 57 s

Niko on its wheeled base, driven around the lab.

Built · Aug – Dec 2022

From a 3D model to a life-size robot

Niko's body is InMoov, Gaël Langevin's open design for a humanoid, whose parts print on a home 3D printer. We built its head, torso, two arms and two five-fingered hands, on a metal stand over a wheeled base. Where InMoov usually runs on Arduino boards and MyRobotLab, Niko runs from one Raspberry Pi 4 with its own Python code.

The build, in photos

  1. Aug 2022

    The 3D model, made from InMoov’s parts in Blender, its name on the chest.

  2. 16 Aug 2022

    One printed hand: five jointed fingers, pulled by tendons.

  3. 2 Nov 2022

    The head on its frame, its eyes in.

  4. 8 Nov 2022

    The head on the chest, “NIKO” printed on it.

  5. 23 Nov 2022

    From the side: the worm gear that turns the waist.

  6. 24 Nov 2022

    Me fitting an arm at the shoulder.

  7. 4 Dec 2022

    Inside the back: the Raspberry Pi 4 between the two servo boards.

  8. 6 Dec 2022

    Launch day, with a camera on its chest and a screen at its base.

  9. 14 Dec 2022

    Interviewed beside Niko by Bangladesh Television.

Before it was finished

  • Aug 2022

    Niko's 3D model in Blender, its joints posed on the timeline

  • Nov 2022

    A head test: the head turns, and the mouth moves

The year in one figure

Niko in 2022, from the first work to the public launch (from the report)

How it works

One Raspberry Pi, 27 servos, a voice and a camera

One Raspberry Pi 4 runs every program. It moves the servos through two PCA9685 boards, hears and speaks through a USB sound card, and sees through a USB camera on its chest.

Datasheet

Computer
Raspberry Pi 4, running Python, started by systemd
Servo boards
Two PCA9685 boards, 16 channels each, set to 60 Hz
Servos
27: head 1, neck 3, eyes 2, jaw 1, and per arm a shoulder, omoplate, rotation, bicep, wrist and five fingers
Camera
USB, on the chest
Sound
A USB sound card, its microphone and speaker shared by every program
Lights
Listening and speaking lights
Screen
A 7-inch screen at the base
Base
Two motors under the stand

The whole robot

System overview, from the Niko report. Tap to open it large.

Motion

Every move goes through one function, servoMove(n, angle). It takes a channel from 0 to 26 and an angle from 0 to 180°, turns the angle into a pulse, and writes it to the first board for channels below 16 or the second board for the rest. No other code needs to know which board a joint is on.

Each gesture is a short Python function that steps joints through angles with pauses between them; most arm moves step one degree every 0.1 s. The handshake reaches out, turns the wrist, closes the fingers halfway for two seconds, opens them and draws back, in about 18.5 s as programmed.

The servo map: every joint's channel and starting angle
The handshake as programmed: four joints over time
The hand poses in the gesture library

Voice

Niko listens for "Hey Niko" with a Porcupine wake word that runs on the Pi. Then the online assistant transcribes the question, and Niko checks it against its own question bank first. The bank is kept in a YAML file (51 questions in 152 phrasings in the published code), so a new answer is an edit to a file, not to the program. Anything else goes to the assistant.

Its own answers are spoken in its own voice, and while it speaks its jaw moves between 70° and 110°, holding each for 0.2 s.

One spoken exchange, from the wake word to the answer

Vision

To follow a face, the camera finds it with OpenCV’s face detector, and the head turns by one fixed rule: the further the face from the middle of the picture, the further the head turns. The camera is fixed on the chest, so the picture does not move as the head turns.

To greet people, each one is enrolled once: eight photos for training and two for testing. Each face becomes a 128-number encoding, and a k-nearest-neighbour classifier names it. Once a name has been on screen for ten frames, Niko says hello and shakes that person’s hand.

Face recognition: enrolment and training, then live greeting
The head-turn rule: head angle against where the face is

Software

A systemd service starts the voice program at boot and restarts it if it stops. The vision programs and the launch speech start on demand, and share modules for the servos, the gestures, the lights and speech. Secrets, such as the wake word’s access key, are read from a file outside the repository.

The software: the voice program, the programs started on demand, and the modules they share

My part, and the team

Built by Quanta Robotics

Niko was built by Quanta Robotics, the robotics team I co-founded at Comilla University, at the FabLab of the Deputy Commissioner’s Office in Cumilla. With the team, I 3D-printed and assembled its body; I tuned its 27 servos, planned its arm motions as servo sequences, and led its programming; most of the commits in its repository are mine.

The team

  • Jewel Nath
  • Shanjit Mondol
  • Tousif Bin Parves
  • Md Mohiuddin Khan Mahin
  • Anik Chakraborty

Niko was launched in public on 6 December 2022. Six news outlets reported it, and Bangladesh Television interviewed me beside it.

Built on

Each file taken from another project keeps its own licence notice; Niko's own code is under the MIT licence.

What I learned

Open hardware shortens the road

Building on InMoov took a student team from a 3D model to a finished life-size humanoid in about four months, and left our effort for the software that makes it useful.

One small computer is enough

A single Raspberry Pi ran the voice, the vision and 27 servos. The PCA9685 boards hold each servo’s pulse on their own, so the Pi only sends new angles.

A local answer bank keeps it in character

An online assistant answers general questions well, but knows nothing about the robot. Matching Niko’s own questions first let it answer about itself and the FabLab in its own voice.

The report

Read the full design

15 pages and 13 diagrams, checked against the code, its history and the dated photos and videos, with the team as co-authors.

Technical report · v1.1 · Zenodo · Oct 2026

Robot Niko: A Life-Size 3D-Printed Humanoid Assistant on a Raspberry Pi — System Design and Build

J. Nath, S. Mondol, T. B. Parves, M. M. K. Mahin, A. Chakraborty

DOI 10.5281/zenodo.23719561

Next: EnviroBot

A rover with a six-servo arm that finds waste with a camera and picks it up, built with my team in 2024.