Agazie Robotics
Get in touch
Research / Agazie
Exploring the frontiers

Research

Early-stage robotics research becomes meaningful when it can survive contact with the real world. We focus on the bridge between both.

Concept render of the Agazie humanoid platform
Humanoid locomotion

Gait optimisation

Balance, posture, joint coordination and efficient motion planning.

Blue particle field forming a face, used as a stand-in for perception research
Perception

Vision pipeline

Detection, recognition, depth and mapping for grounded robotic understanding.

Concept render of Agazie DeskBot
Embodied AI

Interactive robotics

Exploring how perception, language and movement work together in useful machines.

Live Perception Sandbox // Edge Vision On-device only. Embeddings stay in localStorage on this device. Forget me deletes them. Nothing is uploaded.

Interactive Research Sandbox: DeskBot Real-Time Face Perception

A live camera demo of on-device face detection and recognition. YuNet finds each face and five landmarks. SFace turns the aligned crop into a 128-number embedding and compares it with cosine similarity. A face is labelled only when the score passes a threshold and clears the next closest enrolment by a margin. Everyone else stays Unknown.

Engine Standby • Camera Off

DeskBot Optical Subsystem // Standby

Click [Initialize Live Camera Feed] above to run YuNet face detection, five-point alignment, and SFace matching in this browser.

YuNet face detector SFace 128-D embeddings ONNX Runtime WebAssembly
FRAME RATE
-- FPS
Target: 30–60 FPS
INFERENCE LATENCY
-- ms
Edge Compute Time
FACES TRACKED
0 Active
Simultaneous Detections
RECOGNITION CONFIDENCE
-- %
Best SFace cosine similarity
ENROLLED IDENTITIES
0 Active
Stored on this device

When this demo is focused: N Enrol  •  M Cycle mode (FULL / CLEAN / MESH)  •  R Reload saved embeddings  •  Esc Cancel
YuNet (MIT) and SFace int8 (Apache-2.0) run locally with ONNX Runtime WebAssembly. Face images are not stored or uploaded.
Simulation

Fail safely before hardware.

Simulation and controlled testing let us iterate on dynamics, perception and planning before every experiment reaches physical hardware.

Field tests

Prove it outside the lab.

The end point for our research is useful behaviour in messy environments — not a benchmark score in isolation.

Engineering log

Day 1 hardware test: Raspberry Pi 5 and ultra-wide vision

A log for the planned DeskBot perception test on a Raspberry Pi 5 with an IMX219 160° camera. Results are not published yet.

Read the hardware test log →