What it is
Camera-observed UAV pursuit with inspectable decisions.
What it does: Depth and a MuJoCo detector build relative target track and free-space text. A local decision model chooses left, right, climb, descend, forward or simulated fire. A browser cockpit shows probabilities and archives episodes.
Why it belongs in the showcase: A robotics-style action tick with full decision traces, open weights, and a clear observation boundary.
Caveats: Local open model, not hosted TypeSafe Jev; single-pair baseline only.
Evidence
Public Apache-2.0 repo (5 GitHub stars as of intake). Ships a one-pursuer / one-evader baseline in a textured city, RGB/depth HUD, six explicit actions, and JSON/JSONL episode logs. Decision path uses NeoHorse-Jev-4B (Hugging Face) over a local System One-compatible endpoint; Jev receives structured text, not raw RGB/depth tensors. Validated runtime noted as Linux, Python 3.12, MuJoCo 3.14.0.
Caveats
- Boundary: this uses a local NeoHorse Jev-style decision model, not TypeSafe hosted Jev
- Multi-UAV coordination is planned; the current release is one pursuer and one evader
- No published pursuit success-rate benchmark in the README at intake; episodes are for inspection and future training
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.