feat(perception): qualify lidar evidence before models

This commit is contained in:
DCCONSTRUCTIONS
2026-07-25 01:00:26 +03:00
parent 37c24b5fa8
commit 3f549f91f1
8 changed files with 1185 additions and 0 deletions
@@ -0,0 +1,55 @@
# ADR 0018: qualify LiDAR evidence before selecting accelerated models
Date: 2026-07-25
Status: accepted
## Context
Mission Core preserves K1 point clouds, poses, cameras and factory calibration
and already runs an external NVIDIA worker for camera perception. The next
candidate components include PointPillars, CenterPoint, Patchwork++, nvblox and
several LiDAR SLAM systems.
Those components do not share the same input assumptions. The current K1
`lio_pcl` source is a vendor-mapped increment in the canonical `map` frame with intensity and a
separate best-effort pose. It is not an unregistered sweep and exposes no
admitted ring, per-point time or IMU stream. The accepted E10 replay pack also
drops intensity.
Installing a component without expressing those facts would let runtime success
be mistaken for geometric validity.
## Decision
1. Mission Core owns a versioned, provider-neutral
`missioncore.lidar-evidence-profile/v1`.
2. Each processing stage receives a deterministic readiness result with
explicit blockers.
3. Existing content-addressed replay schemas are immutable. A lossless replay
improvement is a new version.
4. The first detector seam consumes sensor-frame XYZI. Map-frame K1 points are
converted only with the bound `T_map_from_lidar` pose, and verified uint8
intensity is normalized to `[0, 1]`.
5. Missing ring, point time, scan geometry or IMU evidence remains missing. It
is never synthesized to satisfy a model.
6. Accelerated models remain external worker providers. React consumes reports,
overlays and world-state products; it does not host CUDA/TensorRT/ROS 2
compute.
7. All LiDAR results remain diagnostic/shadow until separate navigation and
safety gates pass.
## Consequences
- PointPillars integration is blocked for replay pack v1 and degraded for the
current live vendor-mapped source.
- Nvblox is blocked until the LiDAR scan geometry and timing/pose contract are
admitted.
- LiDAR odometry and LiDAR-inertial SLAM are blocked for current K1 evidence.
- L1 lossless replay and scanner-quality telemetry precede new model installs.
- Future scanners, simulation providers and datasets can use the same
readiness contract without being forced into K1-specific code.
## References
The detailed product rationale, market review, gates and sequence are in
`docs/13_LIDAR_WORKER_PRODUCT_AND_ROADMAP.md`.