Kiko is a social robot combining custom SLAM and an expression engine. The SLAM implementation is designed for model hot-swap as vision models improve, and Kiko improves with them. Built entirely in Rust.

Status: Early development. Stereo visual odometry with local bundle adjustment is working.

Structure

kiko/
├── crates/
│   ├── kiko-slam/    # SLAM, visual odometry, feature detection
│   └── oak-sys/      # OAK-D camera FFI bindings
├── comms/
│   ├── robot-server/ # Communication hub (UDP, Serial, HTTP)
│   └── desktop-client/ # Tauri control UI
└── embedded/         # STM32F446 firmware

Quick Start

Record a dataset (requires OAK-D):

cargo run -p kiko-slam --features record -- record recordings/<name>

Live match visualization (requires OAK-D + Rerun viewer):

cargo run -p kiko-slam --features record -- live

Visualize stereo matches on a dataset:

cargo run -p kiko-slam -- viz recordings/<name>

Run visual odometry on a dataset (Rerun viewer):

cargo run -p kiko-slam -- viz --odometry recordings/<name>

Benchmark a dataset:

cargo run -p kiko-slam -- bench recordings/<name>

Config (flags or env)

Inference:

• --downscale / KIKO_DOWNSCALE — input downscale factor (1, 2, or 4)
• --max-keypoints / KIKO_MAX_KEYPOINTS — max keypoints per frame (default 1024)
• --backend / KIKO_BACKEND — inference backend for both models
• --sp-backend / KIKO_SUPERPOINT_BACKEND — SuperPoint backend override
• --lg-backend / KIKO_LIGHTGLUE_BACKEND — LightGlue backend override
• --sp-model / KIKO_SUPERPOINT_MODEL — custom SuperPoint ONNX path
• --lg-model / KIKO_LIGHTGLUE_MODEL — custom LightGlue ONNX path

Visualization:

• --rerun-decimation / KIKO_RERUN_DECIMATION — image decimation for Rerun
• --odometry / KIKO_VIZ_ODOMETRY — enable visual odometry in viz mode
• --allow-unrectified / KIKO_ALLOW_UNRECTIFIED — skip rectification check

Bundle adjustment:

• KIKO_BA_WINDOW — sliding window size (default 10)
• KIKO_BA_ITERS — Gauss-Newton iterations (default 6)
• KIKO_BA_MIN_OBS — minimum observations per frame
• KIKO_BA_HUBER_PX — Huber robust cost threshold in pixels
• KIKO_BA_DAMPING — LM damping factor
• KIKO_BA_MOTION_WEIGHT — motion prior weight (0 to disable)

Keyframe policy:

• KIKO_KF_MIN_INLIERS — inlier count below which a keyframe is forced
• KIKO_KF_PARALLAX_PX — median parallax (px) above which a keyframe is created
• KIKO_KF_MIN_COVISIBILITY — covisibility ratio below which a keyframe is created

Models

Default model paths are resolved under crates/kiko-slam/models/:

• sp.onnx (SuperPoint)
• lg.onnx (LightGlue)

Override with --sp-model / --lg-model or KIKO_SUPERPOINT_MODEL / KIKO_LIGHTGLUE_MODEL.

Roadmap

• Quick dataset recording
• Live match visualisation in Rerun
• Unified CLI (record/live/viz/bench)
• Pipeline benchmarks + per-stage timing
• Stereo triangulation
• Frame-to-keyframe tracking (PnP + RANSAC)
• Local bundle adjustment (sliding window Gauss-Newton)
• Parallax + covisibility keyframe policy
• Keyframe database + map point management
• Covisibility graph
• Place recognition (NetVLAD / CosPlace ONNX)
• Loop closure (Sim3 + pose graph correction)
• Global bundle adjustment
• Dense mapping via nvblox (TSDF / ESDF on Jetson)