Type-generic math
Header-only C++20 numerics — Lie groups, non-linear least squares, camera
models — templated over the scalar type, so the same code runs in double,
float, or Universal posits for mixed-precision study.
Cortex
A layered Rust + C++20 perception stack where algorithms never see middleware and middleware never sees hardware.
What Cortex is
Section titled “What Cortex is”Cortex is the perception core of the Branes Embodied AI framework: a real-time visual-inertial odometry (VIO) and SLAM stack designed to run on a KPU — a spatial-dataflow accelerator — with a host CPU acting as a thin resource broker.
The defining constraint is strict layering: a Rust Resource Manager owns the hardware, header-only C++20 math sits on top, the operator SDK builds algorithms from that math, and middleware (Zenoh/ROS) lives only in the daemons. Algorithms never see middleware; middleware never sees hardware.
MSCKF visual-inertial odometry
A complete Multi-State Constraint Kalman Filter: IMU preintegration and initialization, a sliding window of cloned poses, and null-space camera updates — clean-room from the papers, no GPL source contact.
Hardware behind a HAL
A Rust MemoryProvider trait abstracts the KPU. The same binary carries a
SITL (host-emulated) and a KPU backend, selected at construction by build
target — the C++ above never branches on hardware.
Validated, not just built
ATE/RPE accuracy metrics, a per-frame latency budget, an EuRoC replay harness, and ≥85% Rust coverage gates — efficiency is a functional requirement, enforced in CI.
The stack at a glance
Section titled “The stack at a glance”| Layer | Path | Language | May depend on |
|---|---|---|---|
| Resource Manager | core/ | Rust | OS, kernel drivers, cxx |
| Math | math/ | header-only C++20 | Universal, MTL5 |
| Operator SDK | sdk/ | C++20 | math/, core/ via cxx, cv/ |
| CV front end | cv/ | C++20 | (image / FAST / KLT / pyramid) |
| Daemons | daemons/ | C++ executables | sdk/, Zenoh, yaml-cpp |
Where to go next
Section titled “Where to go next”- Introduction — the problem, the topology, and the design philosophy.
- Architecture — the Host-Broker / Thin-Client topology and the layering invariants.
- Visual-Inertial Odometry — the MSCKF pipeline, end to end.
- Benchmarks & Validation — what was measured, and how.
- Algorithms & Provenance — the clean-room citation table.
Status: Phases 0–3 complete (foundation, Rust core, math layer, VIO). This site is the consolidated record of what was built, the algorithms used, and the performance measured.