Independent developer project · not affiliated with or endorsed by BrainChip, Inc.docs checked 26 SEP 2026
DocsProjectsDemosLab NotesResourcesAbout
PROJECT / ACTIVE RESEARCH BUILD

VORSA-M3

A neuromorphic vision pipeline aimed at detecting freely defined classes, different shapes and sizes, and local triple overlaps in a real production workcell. The page documents the engineering target and the validation path — not a finished benchmark.

EXPERIMENTAL

Current architecture target

ACCELERATORS4 × AKD1500
HOSTRaspberry Pi
VISION2 × Pi Camera 3
WORKLOADSDetection + segmentation
OBJECTSVariable classes / sizes
OVERLAPM1 / M2 / M3 local cases
WHAT WE ARE TRYING TO PROVE

Useful under production constraints.

The meaningful result is not “the model runs.” We need repeatable detection quality, bounded latency, stable multi-device operation and predictable behavior on overlap cases.

WHAT WE WILL NOT CLAIM EARLY

No benchmark without a test harness.

Latency, throughput, power and accuracy values stay off the page until the hardware configuration, dataset slice and measurement method are recorded with them.

VALIDATION ROADMAP

Build confidence in layers.

A four-chip pipeline only makes sense after single-device behavior and model mapping are stable.

PHASE 01Single AKD1500 discovery + stable driverBASELINE
PHASE 02Virtual + physical model mapping parityVALIDATE
PHASE 03Single-device detector benchmarkVALIDATE
PHASE 04Segmentation path / model feasibilityEXPERIMENT
PHASE 05Multi-device assignment and switchingEXPERIMENT
PHASE 06Camera → inference → workcell integrationEXPERIMENT