ruvector-mmwave

Streaming parser for the Seeed MR60BHA2 60 GHz radar's UART protocol. Pure-Rust, no_std-compatible, zero-allocation hot path.

Status: library, 11 unit tests + proptest fuzz suite passing. Shared between the host-side ruvector-mmwave-bridge (parses serial input → emits NL events → cluster embed RPC) and the iter-115 firmware self-test that runs on the radar's MCU directly.

Why a separate crate

ADR-178 Gap H: keeping the parser separate from ruvector-hailo-cluster (the bridge's home crate) means a regression in either side surfaces independently. The parser is byte-for-byte deterministic against fuzzed inputs; the bridge layers transport, TLS, fingerprinting on top.

Wire format (Seeed MR60BHA2 v0.3)

8-byte header  | variable payload | trailing checksum
[0x01]         | <up to 64 bytes>  | invert_xor(payload)
[frame_id_hi]
[frame_id_lo]
[length_hi]    ← 16-bit big-endian payload length
[length_lo]
[type_hi]      ← 16-bit big-endian frame type
[type_lo]
[invert_xor of 7 prior bytes]

Frame types currently parsed:

frame_type	meaning	payload shape
0x0A05	breathing rate	[bpm: u8]
0x0A06	heart rate	[bpm: u8]
0x0A14	nearest target distance	[cm: u16 BE]
0x0F09	presence flag	[present: bool]
anything else	Event::Unknown { frame_type, payload_len }	(iter 249) payload_len is u16

API surface

use ruvector_mmwave::{Event, Mr60Parser};

let mut p = Mr60Parser::new();
let frame: &[u8] = /* 60 bytes from /dev/ttyUSB0 */;
p.feed_slice(frame, |ev| match ev {
    Event::Breathing { bpm } => println!("breathing {} bpm", bpm),
    Event::HeartRate { bpm } => println!("heart rate {} bpm", bpm),
    Event::Distance { cm } => println!("distance {} cm", cm),
    Event::Presence { present } => println!("present={}", present),
    Event::Unknown { frame_type, payload_len } => {
        eprintln!("unknown frame 0x{:04x} len={}", frame_type, payload_len);
    }
    Event::ChecksumError => eprintln!("dropped frame, parser resynced"),
    Event::Resync { skipped } => eprintln!("desync, dropped {} bytes", skipped),
});

The closure signature is FnMut(Event); the parser invokes it zero-or-more times per byte fed. State machine resyncs cleanly on checksum failure or unexpected SOF — no manual reset needed.

Benchmarks

Run cargo bench -p ruvector-mmwave for the full criterion sweep. Steady-state on cognitum-v0 (Pi 5):

• ~3.2 GB/s feed rate on feed_slice with all-recognized frames (most expensive event type)
• ~7.1 GB/s on the no-event-emitted path (waiting for SOF)
• Zero allocations per byte fed — the buffer is fixed at 64 bytes, see MAX_PAYLOAD const.

For typical UART rates (115200 baud → ~14 KB/s), the parser cost is < 0.001% of one core.

Property tests

tests/tokenizer_proptest.rs (proptest v1) feeds:

• arbitrary-length byte strings to verify the parser never panics
• frames with corrupted checksums to verify clean Resync
• frames with valid headers but truncated payloads to verify the state machine waits without emitting

See also

• crates/ruvector-hailo-cluster/src/bin/mmwave-bridge.rs — the host-side bridge that consumes this parser and posts NL events to the hailo-backend cluster.
• examples/esp32-mmwave-sensor/ — firmware-side use of this parser on an ESP32 paired to the radar over UART.
• ADR-063 — original mmwave integration design.
• ADR-178 Gap H — rationale for the separate crate boundary.