[AZ-333] C1 VINS-Mono strategy — research-only comparative VIO

VinsMonoStrategy: Python facade conforming to AZ-331 Protocol; mirrors
the AZ-332 OKVIS2 facade so the AZ-331 factory + IT-12 comparative
harness can treat both as drop-in substitutable. Native binding is a
pybind11 skeleton compiled behind BUILD_VINS_MONO=ON (default OFF for
airborne / operator-tooling / replay-cli per module-layout.md
Build-Time Exclusion Map). Real vins_estimator wiring is the Tier-2
follow-up.

VinsMonoConfig added to c1_vio/config.py with sliding-window /
feature-tracker / marginalisation / opt-iteration knobs plus
__post_init__ validation; exported through the package __init__.

cpp/vins_mono/CMakeLists.txt replaces the AZ-263 placeholder with full
pybind11 wiring: Risk-1 mitigation forces VINS_MONO_USE_ROS=OFF;
Risk-2 mitigation links Eigen from the same cpp/_third_party/eigen pin
as OKVIS2; Risk-3 mitigation enforces BUILD_VINS_MONO=OFF in
deployment binaries via the gate at the top of the file.

Tests: 17 new in test_vins_mono_strategy.py (15 pass + 2 tier2 skip);
fake_vins_mono_binding fixture added to conftest.py mirroring the
fake_okvis2_binding pattern; test_protocol_conformance updated to drop
vins_mono from _STRATEGIES_WITHOUT_PY_MODULE so the existing
parametrised factory tests route through the new strategy.

Focused c1_vio suite: 72 passed, 4 skipped. Full suite: 1788 passed,
1 unrelated pre-existing flake (c12 cold-start perf, env-bound).

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Oleksandr Bezdieniezhnykh
2026-05-14 01:11:09 +03:00
parent 2ce300ddb1
commit 6a5954bdae
13 changed files with 2056 additions and 15 deletions
@@ -25,7 +25,11 @@ from gps_denied_onboard._types.nav import (
VioState,
WarmStartPose,
)
from gps_denied_onboard.components.c1_vio.config import C1VioConfig, Okvis2Config
from gps_denied_onboard.components.c1_vio.config import (
C1VioConfig,
Okvis2Config,
VinsMonoConfig,
)
from gps_denied_onboard.components.c1_vio.errors import (
VioDegradedError,
VioError,
@@ -41,6 +45,7 @@ __all__ = [
"C1VioConfig",
"FeatureQuality",
"Okvis2Config",
"VinsMonoConfig",
"VioDegradedError",
"VioError",
"VioFatalError",
@@ -0,0 +1,338 @@
// AZ-333 — pybind11 binding for VINS-Mono (research-only C1 VIO).
//
// Exposes a narrow surface that mirrors what the Python facade
// (`gps_denied_onboard.components.c1_vio.vins_mono.VinsMonoStrategy`)
// needs — NOT the full VINS-Mono estimator API. The surface mirrors
// the AZ-332 OKVIS2 binding 1:1 so the AZ-331 factory can treat both
// strategies as drop-in substitutable for the IT-12 comparative-study
// research binary:
//
// VinsMonoBackend
// ctor(yaml_config: str, camera_intrinsics_3x3: ndarray[float64, 3, 3])
// add_frame(frame_id: str, ts_ns: int, image: ndarray[uint8, H, W, C]) -> bool
// add_imu(ts_ns: int, accel: ndarray[float64, 3], gyro: ndarray[float64, 3]) -> None
// get_latest_output() -> dict | None
// reset(body_T_world: ndarray[float64, 4, 4], velocity: ndarray[float64, 3],
// accel_bias: ndarray[float64, 3], gyro_bias: ndarray[float64, 3]) -> None
// health() -> dict
//
// Frame buffers cross the FFI boundary as `py::array_t<uint8_t,
// c_style|forcecast>` so the camera-ingest path (AZ-265
// LiveCameraFrameSource) can hand off a contiguous numpy array without
// a copy — Risk-2 mitigation per the AZ-333 task spec.
//
// Exception envelope: every VINS-Mono / Ceres / Eigen / std::runtime_error
// inside a binding method is caught and rethrown as one of three
// Python-side exceptions registered via `py::register_exception`. The
// Python facade then rewraps those into the VioError family.
//
// Risk-1 mitigation (ROS leak): this binding compiles against the
// de-ROSified VINS-Mono pin only — `cpp/vins_mono/CMakeLists.txt`
// strips upstream's `roscpp` / `rosbag` deps before the
// `vins_estimator` core is exposed here.
#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>
#include <pybind11/stl.h>
#include <Eigen/Core>
#include <Eigen/Geometry>
#include <array>
#include <cstdint>
#include <memory>
#include <mutex>
#include <optional>
#include <stdexcept>
#include <string>
// VINS-Mono estimator headers. The exact include path is determined by
// the de-ROSified upstream pin's CMake export. The skeleton compiles
// without these headers because the actual `vins_estimator::Estimator`
// wiring lives in _build_estimator() / _drive_estimator(), which today
// STUB and surface a runtime error if invoked. Wiring them in is the
// follow-up task within AZ-333's tier2 deliverable bundle.
//
// #include <vins_estimator/estimator.h>
// #include <vins_estimator/feature_tracker.h>
// #include <vins_estimator/parameters.h>
namespace py = pybind11;
namespace {
// ---------------------------------------------------------------------------
// Exception types — registered as Python-side classes via
// `py::register_exception` in PYBIND11_MODULE below. The Python facade
// catches these and rewraps into the VioError family.
class VinsMonoInitException : public std::runtime_error {
public:
using std::runtime_error::runtime_error;
};
class VinsMonoFatalException : public std::runtime_error {
public:
using std::runtime_error::runtime_error;
};
class VinsMonoOptimizationException : public std::runtime_error {
public:
using std::runtime_error::runtime_error;
};
// ---------------------------------------------------------------------------
// Pose / output struct produced by the estimator step.
struct EstimatorOutput {
std::string frame_id;
Eigen::Matrix4d pose_T_world_body;
Eigen::Matrix<double, 6, 6> pose_covariance_6x6;
Eigen::Vector3d accel_bias;
Eigen::Vector3d gyro_bias;
int tracked_features = 0;
int new_features = 0;
int lost_features = 0;
double mean_parallax = 0.0;
double mre_px = 0.0;
std::int64_t emitted_at_ns = 0;
};
// ---------------------------------------------------------------------------
// Internal estimator state machine — INIT until the SfM bootstrap
// converges (VINS-Mono's `solve_initial` flips), TRACKING during nominal
// optimisation, DEGRADED on feature-count drop, LOST after consecutive
// failed updates.
enum class HealthState : int { Init = 0, Tracking = 1, Degraded = 2, Lost = 3 };
const char* state_to_str(HealthState s) {
switch (s) {
case HealthState::Init:
return "init";
case HealthState::Tracking:
return "tracking";
case HealthState::Degraded:
return "degraded";
case HealthState::Lost:
return "lost";
}
return "init";
}
// ---------------------------------------------------------------------------
// VinsMonoBackend — the C++ surface exposed to Python.
class VinsMonoBackend {
public:
VinsMonoBackend(const std::string& yaml_config,
py::array_t<double, py::array::c_style | py::array::forcecast>
camera_intrinsics_3x3)
: yaml_config_(yaml_config) {
if (camera_intrinsics_3x3.ndim() != 2 ||
camera_intrinsics_3x3.shape(0) != 3 ||
camera_intrinsics_3x3.shape(1) != 3) {
throw VinsMonoInitException(
"VinsMonoBackend: camera_intrinsics_3x3 must be a 3x3 float64 array");
}
auto buf = camera_intrinsics_3x3.unchecked<2>();
for (py::ssize_t i = 0; i < 3; ++i) {
for (py::ssize_t j = 0; j < 3; ++j) {
K_(i, j) = buf(i, j);
}
}
_build_estimator();
}
// Push a nav-camera frame into the estimator.
// Returns true if the estimator produced a new output for this frame
// (caller then calls `get_latest_output()`); false if the frame was
// consumed but did not yield a new output (e.g. dropped as
// non-keyframe by the parallax-driven keyframe selector).
bool add_frame(
const std::string& frame_id, std::int64_t ts_ns,
py::array_t<std::uint8_t,
py::array::c_style | py::array::forcecast> image) {
if (image.ndim() < 2 || image.ndim() > 3) {
throw VinsMonoOptimizationException(
"VinsMonoBackend.add_frame: image must be 2-D (grayscale) or 3-D "
"(HxWxC)");
}
pending_frame_id_ = frame_id;
pending_ts_ns_ = ts_ns;
return _drive_estimator(image);
}
void add_imu(std::int64_t ts_ns,
py::array_t<double,
py::array::c_style | py::array::forcecast> accel,
py::array_t<double,
py::array::c_style | py::array::forcecast> gyro) {
if (accel.size() != 3 || gyro.size() != 3) {
throw VinsMonoOptimizationException(
"VinsMonoBackend.add_imu: accel and gyro must be length-3 float64 "
"arrays");
}
if (ts_ns <= last_imu_ts_ns_) {
throw VinsMonoOptimizationException(
"VinsMonoBackend.add_imu: ts_ns must be strict-monotonic");
}
last_imu_ts_ns_ = ts_ns;
// Real VINS-Mono IMU push lands here once the estimator is wired
// in. For the skeleton we just record the most recent sample — the
// estimator's IMU pre-integration is performed inside
// `vins_estimator::Estimator::processIMU`.
auto a = accel.unchecked<1>();
auto g = gyro.unchecked<1>();
last_accel_ = Eigen::Vector3d(a(0), a(1), a(2));
last_gyro_ = Eigen::Vector3d(g(0), g(1), g(2));
}
std::optional<py::dict> get_latest_output() const {
std::lock_guard<std::mutex> lk(output_mtx_);
if (!latest_output_.has_value()) {
return std::nullopt;
}
const auto& o = *latest_output_;
py::dict d;
d["frame_id"] = o.frame_id;
d["pose_T_world_body"] = py::array_t<double>(
{4, 4}, {sizeof(double) * 4, sizeof(double)},
o.pose_T_world_body.data());
d["pose_covariance_6x6"] = py::array_t<double>(
{6, 6}, {sizeof(double) * 6, sizeof(double)},
o.pose_covariance_6x6.data());
d["accel_bias"] = py::array_t<double>(
{3}, {sizeof(double)}, o.accel_bias.data());
d["gyro_bias"] = py::array_t<double>(
{3}, {sizeof(double)}, o.gyro_bias.data());
d["tracked_features"] = o.tracked_features;
d["new_features"] = o.new_features;
d["lost_features"] = o.lost_features;
d["mean_parallax"] = o.mean_parallax;
d["mre_px"] = o.mre_px;
d["emitted_at_ns"] = o.emitted_at_ns;
return d;
}
void reset(py::array_t<double,
py::array::c_style | py::array::forcecast> body_T_world,
py::array_t<double,
py::array::c_style | py::array::forcecast> velocity,
py::array_t<double,
py::array::c_style | py::array::forcecast> accel_bias,
py::array_t<double,
py::array::c_style | py::array::forcecast> gyro_bias) {
if (body_T_world.ndim() != 2 || body_T_world.shape(0) != 4 ||
body_T_world.shape(1) != 4) {
throw VinsMonoInitException(
"VinsMonoBackend.reset: body_T_world must be a 4x4 float64 array");
}
if (velocity.size() != 3 || accel_bias.size() != 3 ||
gyro_bias.size() != 3) {
throw VinsMonoInitException(
"VinsMonoBackend.reset: velocity / *_bias must be length-3 float64 "
"arrays");
}
auto T = body_T_world.unchecked<2>();
for (py::ssize_t i = 0; i < 4; ++i) {
for (py::ssize_t j = 0; j < 4; ++j) {
seed_body_T_world_(i, j) = T(i, j);
}
}
auto v = velocity.unchecked<1>();
auto ab = accel_bias.unchecked<1>();
auto gb = gyro_bias.unchecked<1>();
seed_velocity_ = Eigen::Vector3d(v(0), v(1), v(2));
seed_accel_bias_ = Eigen::Vector3d(ab(0), ab(1), ab(2));
seed_gyro_bias_ = Eigen::Vector3d(gb(0), gb(1), gb(2));
state_ = HealthState::Init;
consecutive_lost_ = 0;
{
std::lock_guard<std::mutex> lk(output_mtx_);
latest_output_.reset();
}
_build_estimator();
}
py::dict health() const {
py::dict d;
d["state"] = std::string(state_to_str(state_));
d["consecutive_lost"] = consecutive_lost_;
d["bias_norm"] = std::sqrt(
seed_accel_bias_.squaredNorm() + seed_gyro_bias_.squaredNorm());
return d;
}
private:
void _build_estimator() {
// Real wiring: instantiate `vins_estimator::Estimator` from
// `yaml_config_`, attach the output callback that fills
// `latest_output_` under `output_mtx_` whenever
// `processMeasurements` produces a new sliding-window solution.
//
// The skeleton intentionally throws on any actual frame ingest so a
// research binary that loads this binding before AZ-333's estimator
// wiring lands cannot silently report misleading poses.
estimator_built_ = false;
}
bool _drive_estimator(
py::array_t<std::uint8_t,
py::array::c_style | py::array::forcecast> /*image*/) {
if (!estimator_built_) {
// Skeleton path — pybind11 binding compiles and loads but the
// VINS-Mono estimator is not yet wired. Tier-2 follow-up wires it.
throw VinsMonoFatalException(
"VinsMonoBackend: VINS-Mono estimator not yet wired — this "
"binding is the AZ-333 skeleton; tier2 follow-up wires "
"vins_estimator::Estimator + feature_tracker");
}
return false;
}
std::string yaml_config_;
Eigen::Matrix3d K_ = Eigen::Matrix3d::Identity();
Eigen::Matrix4d seed_body_T_world_ = Eigen::Matrix4d::Identity();
Eigen::Vector3d seed_velocity_ = Eigen::Vector3d::Zero();
Eigen::Vector3d seed_accel_bias_ = Eigen::Vector3d::Zero();
Eigen::Vector3d seed_gyro_bias_ = Eigen::Vector3d::Zero();
Eigen::Vector3d last_accel_ = Eigen::Vector3d::Zero();
Eigen::Vector3d last_gyro_ = Eigen::Vector3d::Zero();
HealthState state_ = HealthState::Init;
int consecutive_lost_ = 0;
std::int64_t last_imu_ts_ns_ = -1;
std::string pending_frame_id_;
std::int64_t pending_ts_ns_ = 0;
bool estimator_built_ = false;
mutable std::mutex output_mtx_;
std::optional<EstimatorOutput> latest_output_;
};
} // namespace
PYBIND11_MODULE(vins_mono_binding, m) {
m.doc() =
"VINS-Mono pybind11 binding (AZ-333). Wraps the de-ROSified VINS-Mono "
"estimator core for the Python VinsMonoStrategy facade. Tier2 follow-up "
"wires the real estimator. Research-only — not present in airborne / "
"operator-tooling / replay-cli binaries (BUILD_VINS_MONO=OFF).";
py::register_exception<VinsMonoInitException>(m, "VinsMonoInitException");
py::register_exception<VinsMonoFatalException>(m, "VinsMonoFatalException");
py::register_exception<VinsMonoOptimizationException>(
m, "VinsMonoOptimizationException");
py::class_<VinsMonoBackend>(m, "VinsMonoBackend")
.def(py::init<const std::string&,
py::array_t<double, py::array::c_style | py::array::forcecast>>(),
py::arg("yaml_config"), py::arg("camera_intrinsics_3x3"))
.def("add_frame", &VinsMonoBackend::add_frame, py::arg("frame_id"),
py::arg("ts_ns"), py::arg("image"))
.def("add_imu", &VinsMonoBackend::add_imu, py::arg("ts_ns"),
py::arg("accel"), py::arg("gyro"))
.def("get_latest_output", &VinsMonoBackend::get_latest_output)
.def("reset", &VinsMonoBackend::reset, py::arg("body_T_world"),
py::arg("velocity"), py::arg("accel_bias"), py::arg("gyro_bias"))
.def("health", &VinsMonoBackend::health);
}
@@ -1,4 +1,4 @@
"""C1 VIO strategy config block (AZ-331 + AZ-332).
"""C1 VIO strategy config block (AZ-331 + AZ-332 + AZ-333).
Registered into ``config.components['c1_vio']`` by the package
``__init__.py``. The composition-root factory
@@ -11,6 +11,12 @@ carrying OKVIS2-specific knobs (sliding-window size, parallax-driven
keyframe threshold, RANSAC inlier ratio, max optimisation iterations,
degraded-feature threshold, per-frame debug log). Only consulted when
``strategy == "okvis2"``.
AZ-333 extends with a sibling :class:`VinsMonoConfig` for the
research-only VINS-Mono backend (sliding-window size, feature tracker
thresholds, marginalisation strategy, max optimisation iterations,
degraded-feature threshold, per-frame debug log). Only consulted when
``strategy == "vins_mono"``.
"""
from __future__ import annotations
@@ -24,6 +30,7 @@ __all__ = [
"KNOWN_STRATEGIES",
"C1VioConfig",
"Okvis2Config",
"VinsMonoConfig",
]
KNOWN_STRATEGIES: Final[frozenset[str]] = frozenset({"okvis2", "vins_mono", "klt_ransac"})
@@ -88,6 +95,85 @@ class Okvis2Config:
)
_ALLOWED_VINS_MARGINALISATION: Final[frozenset[str]] = frozenset(
{"old", "second_new"}
)
@dataclass(frozen=True)
class VinsMonoConfig:
"""VINS-Mono-specific knobs (AZ-333; research-only backend).
``sliding_window_size`` is the VINS-Mono optimisation-window size
in keyframes — must be in [10, 20] mirroring D-C5-3's K bound.
``feature_min_tracked`` is the per-frame tracked-feature floor
below which the frontend declares the frame untrackable; default
20 (VINS-Mono ``MIN_DIST`` upstream default surface).
``feature_min_parallax_px`` is the parallax-driven keyframe
selection threshold; default 10.0 px (VINS-Mono upstream default
for 752×480 EuRoC-class fixtures).
``marginalisation_strategy`` selects ``"old"`` (drop the oldest
keyframe and marginalise its prior into the Hessian) or
``"second_new"`` (drop the second-newest, used when the newest is
a non-keyframe). Both are upstream-supported.
``max_optimization_iters`` caps the per-frame Ceres solver
iterations; default 8 (VINS-Mono upstream default; higher than
OKVIS2 because Ceres single-iteration cost is lower).
``degraded_feature_threshold`` is the per-frame tracked-feature
count below which ``health_snapshot`` reports DEGRADED; default 30
(matches ``Okvis2Config`` so cross-strategy comparison is fair).
``per_frame_debug_log`` enables a DEBUG log line per
``process_frame`` — OFF by default.
"""
sliding_window_size: int = 10
feature_min_tracked: int = 20
feature_min_parallax_px: float = 10.0
marginalisation_strategy: str = "old"
max_optimization_iters: int = 8
degraded_feature_threshold: int = 30
per_frame_debug_log: bool = False
def __post_init__(self) -> None:
if not (10 <= self.sliding_window_size <= 20):
raise ConfigError(
"VinsMonoConfig.sliding_window_size must be in [10, 20] "
f"(D-C5-3 budget); got {self.sliding_window_size}"
)
if self.feature_min_tracked < 1:
raise ConfigError(
"VinsMonoConfig.feature_min_tracked must be >= 1; "
f"got {self.feature_min_tracked}"
)
if self.feature_min_parallax_px <= 0.0:
raise ConfigError(
"VinsMonoConfig.feature_min_parallax_px must be > 0; "
f"got {self.feature_min_parallax_px}"
)
if self.marginalisation_strategy not in _ALLOWED_VINS_MARGINALISATION:
raise ConfigError(
"VinsMonoConfig.marginalisation_strategy must be one of "
f"{sorted(_ALLOWED_VINS_MARGINALISATION)}; "
f"got {self.marginalisation_strategy!r}"
)
if self.max_optimization_iters < 1:
raise ConfigError(
"VinsMonoConfig.max_optimization_iters must be >= 1; "
f"got {self.max_optimization_iters}"
)
if self.degraded_feature_threshold < 1:
raise ConfigError(
"VinsMonoConfig.degraded_feature_threshold must be >= 1; "
f"got {self.degraded_feature_threshold}"
)
@dataclass(frozen=True)
class C1VioConfig:
"""Per-component config for C1 VIO.
@@ -106,12 +192,16 @@ class C1VioConfig:
``okvis2`` carries OKVIS2-specific knobs (AZ-332); consulted only
when ``strategy == "okvis2"``.
``vins_mono`` carries VINS-Mono-specific knobs (AZ-333); consulted
only when ``strategy == "vins_mono"``.
"""
strategy: str = "klt_ransac"
lost_frame_threshold: int = 9
warm_start_max_frames: int = 5
okvis2: Okvis2Config = field(default_factory=Okvis2Config)
vins_mono: VinsMonoConfig = field(default_factory=VinsMonoConfig)
def __post_init__(self) -> None:
if self.strategy not in KNOWN_STRATEGIES:
@@ -0,0 +1,518 @@
"""`VinsMonoStrategy` — research-only comparative C1 VIO (AZ-333).
Python facade over the VINS-Mono C++ loosely-coupled sliding-window VIO
core, accessed via the pybind11 binding at
``_native.vins_mono_binding.VinsMonoBackend`` (compiled by
``cpp/vins_mono/CMakeLists.txt``, gated by ``BUILD_VINS_MONO=ON``).
Conforms to the AZ-331 :class:`VioStrategy` Protocol; consumes the
runtime ``Config`` + an :class:`FdrClient`; constructs its other
dependencies (logger, camera calibration) internally from ``config``
so the strategy class matches the composition-root factory shape::
strategy_cls(config: Config, *, fdr_client: FdrClient)
This mirrors :class:`Okvis2Strategy` (AZ-332) deliberately: the AZ-331
factory produces both via the same `(config, *, fdr_client)` shape and
the IT-12 comparative-study harness expects the two to be drop-in
substitutable. Behavioural differences (Ceres vs Levenberg-Marquardt,
loosely-coupled vs tightly-coupled, marginalisation strategy) live
under the binding boundary and are observable only via the latency /
covariance numbers in the Step 9 comparative report — NOT via the
Python surface.
Risk-2 / Risk-3 mitigation: the native binding is imported **lazily
inside the constructor**, not at module top level. Importing this
module with ``BUILD_VINS_MONO=OFF`` (no compiled ``.so``) is safe —
the AZ-331 factory's build-flag gate catches that path before the
constructor runs.
AC mapping (see ``_docs/02_tasks/todo/AZ-333_c1_vins_mono_strategy.md``):
- AC-1 : :meth:`current_strategy_label` returns ``"vins_mono"``.
- AC-2 : :meth:`process_frame` returns :class:`VioOutput` with
``frame_id`` echoed; covariance SPD; ``imu_bias`` non-None.
- AC-3 : all backend / Ceres / Eigen / std::runtime_error rewrap into
:class:`VioError` family with ``__cause__`` chain.
- AC-4 : :meth:`reset_to_warm_start` clears state + seeds hint; second
consecutive call does not raise.
- AC-5 : :meth:`health_snapshot` returns INIT initially, TRACKING after
``warm_start_max_frames`` (default 5) successful frames.
- AC-6 : DEGRADED on feature loss; covariance Frobenius norm strictly
increases; ``process_frame`` still returns :class:`VioOutput` (not raise).
- AC-7 : after ``lost_frame_threshold`` (default 9) consecutive failed
frames, raises :class:`VioFatalError`; state == LOST.
- AC-8 : ``BUILD_VINS_MONO=OFF`` does not load this module (enforced by
AZ-331 factory; covered in
``tests/unit/c1_vio/test_protocol_conformance.py``).
- AC-9 / NFR-perf : tier2 — Jetson + Derkachi-class fixture; tests
marked ``@pytest.mark.tier2``.
- AC-10 : exactly one ``vio.health`` FDR record per state transition;
no spam on steady-state.
"""
from __future__ import annotations
import math
from datetime import datetime, timezone
from typing import TYPE_CHECKING, Any, Final, Literal
import numpy as np
from gps_denied_onboard._types.nav import (
FeatureQuality,
ImuBias,
VioHealth,
VioOutput,
VioState,
)
from gps_denied_onboard.clock.wall_clock import WallClock
from gps_denied_onboard.components.c1_vio.errors import (
VioFatalError,
VioInitializingError,
)
from gps_denied_onboard.fdr_client.records import CURRENT_SCHEMA_VERSION, FdrRecord
from gps_denied_onboard.logging import get_logger
if TYPE_CHECKING:
import numpy.typing as npt
from gps_denied_onboard._types.calibration import CameraCalibration
from gps_denied_onboard._types.nav import (
ImuWindow,
NavCameraFrame,
WarmStartPose,
)
from gps_denied_onboard.clock import Clock
from gps_denied_onboard.components.c1_vio.config import VinsMonoConfig
from gps_denied_onboard.config import Config
from gps_denied_onboard.fdr_client.client import FdrClient
__all__ = ["VinsMonoStrategy"]
_STRATEGY_LABEL: Final[Literal["vins_mono"]] = "vins_mono"
_PRODUCER_ID: Final[str] = "c1_vio.vins_mono"
_LOGGER_COMPONENT: Final[str] = "c1_vio.vins_mono"
def _now_iso() -> str:
return datetime.now(timezone.utc).isoformat()
def _bias_norm(bias: ImuBias) -> float:
"""L2 norm of the concatenated 6-vector ``(accel || gyro)``."""
accel = np.asarray(bias.accel_bias, dtype=np.float64)
gyro = np.asarray(bias.gyro_bias, dtype=np.float64)
return float(math.sqrt(float(np.dot(accel, accel) + np.dot(gyro, gyro))))
def _se3_from_4x4(matrix: npt.NDArray[Any]) -> Any:
"""Build a ``gtsam.Pose3`` from a 4x4 row-major matrix.
Imported lazily so this module can be imported without gtsam in
headless tooling paths (tests + facade-only smoke).
"""
import gtsam
return gtsam.Pose3(np.asarray(matrix, dtype=np.float64))
class VinsMonoStrategy:
"""Research-only :class:`VioStrategy` for IT-12 comparative study (AZ-333).
Constructor matches the AZ-331 composition-root factory shape::
VinsMonoStrategy(config: Config, *, fdr_client: FdrClient)
Other dependencies (calibration, logger, VINS-Mono sub-config) are
resolved internally from ``config``. Per the C1 component
`tests.md` C1-IT-04, the AC-2.2 MRE bound is **exempt** for this
strategy.
Concurrency: single-threaded by Protocol invariant. One instance
per camera-ingest writer thread; concurrent ``process_frame`` calls
are undefined behaviour.
"""
def __init__(
self,
config: Config,
*,
fdr_client: FdrClient,
clock: Clock | None = None,
) -> None:
c1_block = config.components["c1_vio"]
if c1_block.strategy != _STRATEGY_LABEL:
raise VioFatalError(
f"VinsMonoStrategy constructed with config.strategy="
f"{c1_block.strategy!r}; expected {_STRATEGY_LABEL!r}. "
"The AZ-331 factory is the only sanctioned constructor."
)
self._config = config
self._fdr = fdr_client
self._clock: Clock = clock if clock is not None else WallClock()
self._logger = get_logger(_LOGGER_COMPONENT)
self._lost_frame_threshold: int = c1_block.lost_frame_threshold
self._warm_start_max_frames: int = c1_block.warm_start_max_frames
self._vins_cfg: VinsMonoConfig = c1_block.vins_mono
self._calibration: CameraCalibration | None = None
self._frames_since_warmup: int = 0
self._consecutive_lost: int = 0
self._latest_bias: ImuBias = ImuBias(
accel_bias=(0.0, 0.0, 0.0), gyro_bias=(0.0, 0.0, 0.0)
)
self._reported_state: VioState = VioState.INIT
self._last_emitted_state: VioState | None = None
# Lazy import of the native binding — Risk-2 / Risk-3 mitigation.
# Failure here is the BUILD_VINS_MONO=OFF path the AZ-331
# factory's `StrategyNotAvailableError` is meant to prevent; if a
# caller bypasses the factory and reaches this constructor with
# the native lib absent, we surface a fatal init error.
try:
from gps_denied_onboard.components.c1_vio._native import (
vins_mono_binding,
)
except ImportError as exc:
raise VioFatalError(
"VinsMonoStrategy: native binding "
"(gps_denied_onboard.components.c1_vio._native.vins_mono_binding) "
"is not importable. Research binary must be built with "
"BUILD_VINS_MONO=ON; deployment binaries (airborne / "
"operator-tooling / replay-cli) must NOT request strategy="
"'vins_mono'."
) from exc
self._binding_module = vins_mono_binding
self._backend = self._construct_backend()
# ------------------------------------------------------------------
# Public Protocol surface.
def process_frame(
self,
frame: NavCameraFrame,
imu: ImuWindow,
calibration: CameraCalibration,
) -> VioOutput:
"""Hot-path call — one per nav-camera frame.
Steps:
1. Push every IMU sample in the window into the backend; the
strict-monotonic guard lives on the C++ side.
2. Submit the frame.
3. If the backend produced an output, classify health and
build the :class:`VioOutput` DTO.
4. If no output: tick the lost-frame counter; emit a state
transition record if needed.
"""
self._calibration = calibration
frame_id_str = str(frame.frame_id)
emitted_at_ns = self._clock.monotonic_ns()
try:
self._push_imu_window(imu)
produced = self._backend.add_frame(
frame_id_str, _frame_ts_ns(frame), _frame_image(frame)
)
except self._binding_module.VinsMonoInitException as exc:
self._emit_transition(VioState.INIT, frame_id_str)
raise VioInitializingError(
f"VINS-Mono backend reports INIT while processing frame "
f"{frame_id_str!r}: {exc}"
) from exc
except self._binding_module.VinsMonoOptimizationException as exc:
# Treat as a degraded frame: emit no VioOutput from this
# path — callers expect either a VioOutput or a VioError;
# we choose error here so C5 can fall back, matching AC-3.
self._tick_lost(frame_id_str)
if self._reported_state == VioState.LOST:
self._emit_transition(VioState.LOST, frame_id_str)
raise VioFatalError(
f"VINS-Mono backend exhausted lost-frame budget at "
f"{frame_id_str!r}: {exc}"
) from exc
self._emit_transition(self._reported_state, frame_id_str)
raise VioInitializingError(
f"VINS-Mono backend optimisation failure at {frame_id_str!r}: {exc}"
) from exc
except self._binding_module.VinsMonoFatalException as exc:
self._emit_transition(VioState.LOST, frame_id_str)
raise VioFatalError(
f"VINS-Mono backend fatal exception at {frame_id_str!r}: {exc}"
) from exc
except (RuntimeError, ValueError) as exc:
# Catch-all for unmapped backend exceptions. Re-classify as
# fatal — we explicitly forbid raw library exceptions across
# the public boundary.
raise VioFatalError(
f"VINS-Mono backend raised an unmapped exception at "
f"{frame_id_str!r}: {exc}"
) from exc
if not produced:
# Frame consumed but no estimator update yet — INIT path
# while VINS-Mono's SfM bootstrap warms up.
self._emit_transition(VioState.INIT, frame_id_str)
raise VioInitializingError(
f"VinsMonoStrategy: backend has not yet emitted an "
f"estimator update at {frame_id_str!r}"
)
raw = self._backend.get_latest_output()
if raw is None:
raise VioFatalError(
f"VinsMonoStrategy: backend reported a new output for "
f"{frame_id_str!r} but get_latest_output() returned None"
)
vio_output = self._build_vio_output(raw, emitted_at_ns)
self._consecutive_lost = 0
new_state = self._classify_state(vio_output.feature_quality)
if new_state != self._reported_state:
self._reported_state = new_state
self._emit_transition(new_state, frame_id_str)
if new_state in (VioState.INIT, VioState.TRACKING):
self._frames_since_warmup += 1
if self._vins_cfg.per_frame_debug_log:
self._logger.debug(
"vins_mono.process_frame",
extra={
"component": _LOGGER_COMPONENT,
"kind": "vio.tick",
"frame_id": frame_id_str,
"kv": {
"state": self._reported_state.value,
"tracked": vio_output.feature_quality.tracked,
"mre_px": vio_output.feature_quality.mre_px,
"emitted_at_ns": vio_output.emitted_at_ns,
},
},
)
return vio_output
def reset_to_warm_start(self, hint: WarmStartPose) -> None:
"""Destructive re-init from an F8-reboot warm-start hint.
Idempotent across consecutive calls (AC-4) — a second call
without an intervening ``process_frame`` reseats the backend
again without raising.
"""
try:
body_T_world = np.asarray(hint.body_T_world.matrix(), dtype=np.float64)
except AttributeError as exc:
raise VioFatalError(
"VinsMonoStrategy.reset_to_warm_start: hint.body_T_world is "
"not a gtsam.Pose3 (missing .matrix())"
) from exc
velocity = np.asarray(hint.velocity_b, dtype=np.float64)
accel_bias = np.asarray(hint.bias.accel_bias, dtype=np.float64)
gyro_bias = np.asarray(hint.bias.gyro_bias, dtype=np.float64)
try:
self._backend.reset(body_T_world, velocity, accel_bias, gyro_bias)
except self._binding_module.VinsMonoInitException as exc:
raise VioFatalError(
f"VINS-Mono backend rejected warm-start reset: {exc}"
) from exc
except (RuntimeError, ValueError) as exc:
raise VioFatalError(
f"VINS-Mono backend raised an unmapped exception during reset: {exc}"
) from exc
self._latest_bias = hint.bias
self._frames_since_warmup = 0
self._consecutive_lost = 0
self._reported_state = VioState.INIT
self._emit_transition(VioState.INIT, frame_id="")
def health_snapshot(self) -> VioHealth:
"""Most-recent health state — no backend call (cheap)."""
return VioHealth(
state=self._reported_state,
consecutive_lost=self._consecutive_lost,
bias_norm=_bias_norm(self._latest_bias),
)
def current_strategy_label(self) -> Literal["okvis2", "vins_mono", "klt_ransac"]:
return _STRATEGY_LABEL
# ------------------------------------------------------------------
# Internal helpers.
def _construct_backend(self) -> Any:
"""Build the backend from config — calibration path is optional
because the unit-test fake-binding path skips real intrinsics.
Tests inject a fake module at ``sys.modules`` before construction
(see ``tests/unit/c1_vio/conftest.py``); the fake's
``VinsMonoBackend`` accepts whatever this method passes.
"""
K = self._load_camera_intrinsics()
yaml_config = self._render_yaml_config()
try:
return self._binding_module.VinsMonoBackend(yaml_config, K)
except self._binding_module.VinsMonoInitException as exc:
raise VioFatalError(
f"VinsMonoStrategy: backend init failed: {exc}"
) from exc
def _load_camera_intrinsics(self) -> np.ndarray:
"""Load 3x3 camera intrinsics from the calibration path.
Returns the identity matrix when the runtime block has no
path configured — the unit-test path overrides this via the
fake binding's ctor anyway, and a research binary refusing
to start on a missing calibration is preferable to silently
emitting wrong poses (handled by the YAML loader downstream).
"""
path = self._config.runtime.camera_calibration_path
if not path:
return np.eye(3, dtype=np.float64)
try:
import json
with open(path, encoding="utf-8") as fh:
blob = json.load(fh)
except (OSError, ValueError) as exc:
raise VioFatalError(
f"VinsMonoStrategy: failed to load camera calibration from "
f"{path!r}: {exc}"
) from exc
K_raw = blob.get("intrinsics_3x3")
if K_raw is None:
raise VioFatalError(
f"VinsMonoStrategy: calibration file {path!r} is missing the "
"'intrinsics_3x3' field"
)
K = np.asarray(K_raw, dtype=np.float64)
if K.shape != (3, 3):
raise VioFatalError(
f"VinsMonoStrategy: intrinsics_3x3 must be 3x3; got shape {K.shape}"
)
return K
def _render_yaml_config(self) -> str:
"""Render the VinsMonoConfig sub-block into a VINS-Mono YAML snippet.
VINS-Mono reads a YAML config string at construction. Only the
knobs AZ-333 exposes are rendered; VINS-Mono-internal defaults
cover the rest.
"""
cfg = self._vins_cfg
return (
"# AZ-333 — generated VINS-Mono config (see VinsMonoConfig in c1_vio/config.py)\n"
f"sliding_window_size: {cfg.sliding_window_size}\n"
f"feature_min_tracked: {cfg.feature_min_tracked}\n"
f"feature_min_parallax_px: {cfg.feature_min_parallax_px}\n"
f"marginalisation_strategy: {cfg.marginalisation_strategy}\n"
f"max_optimization_iters: {cfg.max_optimization_iters}\n"
)
def _push_imu_window(self, imu: ImuWindow) -> None:
for sample in imu.samples:
self._backend.add_imu(
sample.ts_ns,
np.asarray(sample.accel_xyz, dtype=np.float64),
np.asarray(sample.gyro_xyz, dtype=np.float64),
)
def _build_vio_output(self, raw: dict[str, Any], emitted_at_ns: int) -> VioOutput:
try:
pose = _se3_from_4x4(raw["pose_T_world_body"])
cov = np.asarray(raw["pose_covariance_6x6"], dtype=np.float64)
bias = ImuBias(
accel_bias=tuple(float(x) for x in raw["accel_bias"]), # type: ignore[arg-type]
gyro_bias=tuple(float(x) for x in raw["gyro_bias"]), # type: ignore[arg-type]
)
feature_quality = FeatureQuality(
tracked=int(raw["tracked_features"]),
new=int(raw["new_features"]),
lost=int(raw["lost_features"]),
mean_parallax=float(raw["mean_parallax"]),
mre_px=float(raw["mre_px"]),
)
backend_ts = int(raw.get("emitted_at_ns") or emitted_at_ns)
except (KeyError, TypeError, ValueError) as exc:
raise VioFatalError(
f"VinsMonoStrategy: backend output is malformed: {exc}"
) from exc
if cov.shape != (6, 6):
raise VioFatalError(
f"VinsMonoStrategy: pose_covariance_6x6 has shape {cov.shape}; "
"expected (6, 6)"
)
self._latest_bias = bias
return VioOutput(
frame_id=raw["frame_id"],
relative_pose_T=pose,
pose_covariance_6x6=cov,
imu_bias=bias,
feature_quality=feature_quality,
emitted_at_ns=backend_ts,
)
def _classify_state(self, fq: FeatureQuality) -> VioState:
if self._reported_state == VioState.INIT and (
self._frames_since_warmup + 1 < self._warm_start_max_frames
):
return VioState.INIT
if fq.tracked < self._vins_cfg.degraded_feature_threshold:
return VioState.DEGRADED
return VioState.TRACKING
def _tick_lost(self, frame_id: str) -> None:
self._consecutive_lost += 1
if self._consecutive_lost >= self._lost_frame_threshold:
self._reported_state = VioState.LOST
elif self._reported_state == VioState.TRACKING:
self._reported_state = VioState.DEGRADED
def _emit_transition(self, new_state: VioState, frame_id: str) -> None:
if self._last_emitted_state == new_state:
return
self._last_emitted_state = new_state
record = FdrRecord(
schema_version=CURRENT_SCHEMA_VERSION,
ts=_now_iso(),
producer_id=_PRODUCER_ID,
kind="vio.health",
payload={
"state": new_state.value,
"consecutive_lost": self._consecutive_lost,
"bias_norm": _bias_norm(self._latest_bias),
"strategy_label": _STRATEGY_LABEL,
"frame_id": frame_id,
},
)
self._fdr.enqueue(record)
def _frame_ts_ns(frame: NavCameraFrame) -> int:
"""Convert ``NavCameraFrame.timestamp`` to monotonic-ns.
Uses the datetime's UTC epoch nanoseconds so the value is
monotonically increasing across frames (frame source guarantees
strictly increasing timestamps per the FrameSource contract).
"""
return int(frame.timestamp.timestamp() * 1e9)
def _frame_image(frame: NavCameraFrame) -> np.ndarray:
"""Coerce the frame's image into a contiguous uint8 ndarray."""
arr = np.ascontiguousarray(frame.image, dtype=np.uint8)
if arr.ndim < 2 or arr.ndim > 3:
raise VioFatalError(
f"VinsMonoStrategy: NavCameraFrame.image must be 2-D or 3-D; "
f"got {arr.ndim}-D"
)
return arr