tpt-perception

A pure-Rust, AI-native 3D spatial perception and sensor fusion library: point cloud processing, SLAM, multi-sensor fusion, geometric feature extraction, and spatial registration — with no C/C++ FFI and no external algorithm libraries.

Part of the tpt-* family. See spec.txt for the full design specification and todo.md for the phased build tracker.

Design principles

• Zero external dependencies for algorithms. ICP, NDT, voxelization, feature detection, filters — all implemented from scratch in pure Rust, built on the in-house tpt-math numeric substrate.
• Strict licensing. MIT OR Apache-2.0 throughout; deny.toml blocks Apache-2.0-only wrap targets.
• Type-state spatial safety. Coordinate frames are phantom type parameters: transforming a Point3D<Sensor<Lidar>> with a Sensor<Lidar> → World isometry compiles; mixing frames does not.
• Native formal verification. Kani bounded model checking harnesses for critical spatial algorithms, property-based tests (proptest) for statistical geometric invariants, documented mathematical contracts.
• no_std + alloc core. All algorithm crates compile without std.
• AI-native. Feature descriptors and cost functions expose hooks for learned (autodiff-ready) perception.

Crates

Crate	Purpose
tpt-percept-core	3D geometry, frame-typed points/transforms, twists, small dense linear algebra
tpt-percept-cloud	Point clouds: voxelization, filtering, k-d tree, octree, R-tree, streaming
tpt-percept-features	Normals, curvature, PCA, Harris 3D, edges, RANSAC planes, FPFH, SHOT, learned-descriptor hooks
tpt-percept-register	ICP (point-to-point / point-to-plane), NDT, feature + RANSAC, branch-and-bound, 3D Hough
tpt-percept-slam	Visual & LiDAR odometry, loop closure, pose graph optimization
tpt-percept-fusion	Temporal alignment, EKF/UKF, extrinsic calibration, covariance intersection, robust fusion
tpt-percept-map	Occupancy grids (2D/3D log-odds), TSDF, signed distance fields, semantic maps
tpt-percept-verify	proptest strategies, geometric invariant checks, Kani proof harnesses
tpt-perception	Feature-gated umbrella crate re-exporting everything above

Quick start

[dependencies]
tpt-perception = "0.1"          # umbrella, all features
# or pick crates à la carte:
tpt-percept-register = "0.1"

use tpt_perception::prelude::*;

// Frames are types, not comments: this does not compile —
// let p_world: Point3D<World> = sensor_point.transform_with(world_to_camera);

Building and verifying

cargo build --workspace
cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --check
cargo deny check            # license/advisory hygiene
cargo build -p tpt-percept-core --no-default-features --features alloc   # no_std gate
# Kani proofs (requires cargo-kani): cargo kani -p tpt-percept-verify

Status

Phase 4 complete — see todo.md for the itemized checklist. Open integration checkpoints (external tpt-systems-optimisation, tpt-gis, tpt-control repos) are tracked at the top of todo.md.

License

Dual-licensed under MIT OR Apache-2.0. © TPT Solutions.