ASTRA 3.0 — CL1 ↔ Unreal Engine, corrected build

Closed-loop bridge between Cortical Labs' CL1 (or a compatible substrate) and Unreal Engine, with an ASTRA MCP supervision path. This build applies every correction from the review cycle and ships a runnable Phase-1 core plus the continuous 3D-navigation decoder.

Honest status. Only what can run, runs. Phase 1 (the UDP bridge, the substrate simulator, the wire protocol, the safety envelope) is implemented and tested here — python3 bridge/bridge.py --selftest passes, and the test suite is green (13/13). Phases 2–4 of the wider ASTRA program (unified IRB router, LLM orchestration, comparative benchmarking) are not reproduced in this package; see PHASES.md for their real status. This repository does not vendor the full ASTRA v2.9 TypeScript MCP server.

What's inside

Related MCP server: UE5 MCP Server

Quick start (no hardware)

# 1. (optional) deps — the Phase-1 core needs only the Python stdlib
pip install -r requirements.txt      # pyyaml (optional), pytest (optional)

# 2. validate the whole pipe in-process
python3 bridge/bridge.py --selftest

# 3. run the substrate over UDP and stream spikes to a UE receiver on :12345
python3 bridge/bridge.py --organoid --ue-host 127.0.0.1 --ue-port 12345 \
        --control-listen-port 12346 --random-seed 42 --duration 5

# 4. run the tests
python3 -m unittest discover -s tests -p 'test_all.py'   # or: pytest tests/ -q

The 3D-navigation fix (headline correction)

The documented decoder pooled 64 channels into 4 horizontal bands (band = channel/16) and took an argmax → one discrete state among Up/Down/Left/Right. That is the wrong abstraction for a 3D view: it discards magnitude (no diagonals, no variable speed), and the horizontal bands carry no column (X) information even though Left/Right is an X distinction. UCl1AxisDecoder replaces it with a rate-weighted centroid over the true 8×8 grid, on baseline-subtracted rates, giving a smooth analog stick + a forward-velocity magnitude — the "axis mappings" read-out of Assembloid Agency §3.4. Full rationale in docs/OPTIMISATION-VUE-3D.md.

Verified references

• CL API — arXiv:2602.11632 (Hogan, Doherty, Khoo, Zhou, Salib, Stewart, Lawson, Loeffler, Kagan, 2026; CC BY-NC-SA 4.0). Hard limit used here: 200 Hz/channel.

• Assembloid Agency — Leung & Loewith, NeurIPS 2025 Creative AI Track. The §3.4 "spike-to-axis" read-out and §3.4 proximity→frequency encoding are implemented.

License

MIT for this package — with a third-party caveat: the Cortical Labs CL SDK (the cl module) is CC BY-NC 4.0 (non-commercial), not on PyPI, and is not bundled here. It is only needed to drive real hardware; the simulator and selftest need none of it. See LICENSE.

Assistance Multi IA · Assistant-Multi-AI@proton.me