Skip to content

Boot Sequence

Jump to bottom
OppaAI edited this page Apr 30, 2026 · 1 revision

Boot Sequence

GRACE requires four services running in order before she's ready. Each step depends on the previous one.

Prerequisites

  • Jetson Orin Nano Super with JetPack 6.2.2
  • ROS2 Humble sourced in your shell
  • vLLM Docker image pulled (embedl/vllm:latest-jetson-orin-flashhead)
  • Repository cloned to ~/AGi

Step-by-Step

1. Clone and build (first time only)

git clone https://github.com/OppaAI/AGi ~/AGi
cd ~/AGi/AuRoRA
 
# Install ROS2 dependencies
rosdep install --from-paths src --ignore-src -r -y
 
# Install Python dependencies
pip3 install -r requirements.txt --break-system-packages
 
# Build the SCS package
colcon build --packages-select scs

2. Source the workspace

Every new terminal needs this:

cd ~/AGi/AuRoRA
source install/setup.bash

3. Start Cosmos vLLM

bash launch/cosmos.sh

Wait for this line before continuing — it takes ~3 minutes on cold start:

INFO:     Application startup complete.

Cosmos runs on http://localhost:8000 and serves the OpenAI-compatible /v1/chat/completions endpoint that CNC streams from.

4. Start GRACE (CNC node)

In a new terminal (with workspace sourced):

ros2 run scs cnc

You should see:

============================================================
🧠 CNC — Central Neural Core starting…
============================================================
   [Working Memory Cortex]  ONLINE ✅ — 7±2 PMT slots | N chunks allocated
⏳ Activating Encoding Engine (BAAI/bge-small-en-v1.5)…
✅ Encoding Engine activated
🔄 EMC encoding cycle started
✅ EMC initialized → /home/<user>/.aurora/neural_gateway/engram_complex/emc.db
✅ Memory Coordination Core Activated
✅ Subscribed  : /cns/neural_input
✅ Publishing  : /gce/response
✅ vLLM        : http://localhost:8000
✅ Model       : embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
============================================================
🌸 GRACE is ready
============================================================

5. Start rosbridge

In a new terminal:

ros2 launch rosbridge_server rosbridge_websocket_launch.xml

rosbridge exposes the ROS2 topics over WebSocket so the GUI can communicate with GRACE.

6. Serve the GUI

python3 -m http.server 9413 --directory ~/AGi/AuRoRA/src/scs/scs

Open in a browser on any device on the same network:

http://<jetson-ip>:9413/AGi.html

Terminal Layout

Running GRACE comfortably needs four terminals side by side:

Terminal Command Purpose
T1 bash launch/cosmos.sh Cosmos vLLM inference server
T2 ros2 run scs cnc GRACE — Central Neural Core
T3 ros2 launch rosbridge_server rosbridge_websocket_launch.xml WebSocket bridge to GUI
T4 python3 -m http.server 9413 --directory src/scs/scs GUI file server

Shutdown

Shut down in reverse order — CNC first, Cosmos last.

  1. Ctrl+C in T2 (CNC) — triggers graceful destroy_node(): closes MCC, drains encoding cycle, releases SQLite connection
  2. Ctrl+C in T3 (rosbridge)
  3. Ctrl+C in T4 (GUI server)
  4. Ctrl+C in T1 (Cosmos) — or docker stop if running in Docker CNC's shutdown is graceful — the encoding cycle gets a stop signal, waits up to 3 seconds for a clean exit, then the engram connection is released. Don't kill the process hard mid-conversation if you can avoid it; unencoded PMTs in the binding stream are protected by the staging table and will be recovered on next boot.

Crash Recovery

If CNC crashes mid-session, any PMTs that were in the binding stream but not yet encoded are safe. On next boot, CNC checks emc_staging and recovers them back into the binding stream before the encoding cycle starts:

⚡ EMC recovered N unencoded episode(s) from engram → binding stream

Working memory (WMC) is in-RAM only — it does not persist across restarts. Episodic memory (EMC) is permanent.

Common Issues

Cosmos not ready — CNC times out on first message Wait for Application startup complete in the Cosmos terminal before sending any messages. CNC will log a timeout error and GRACE will respond with an error chunk — just try again once Cosmos is up.

Encoding Engine offline warning at startup sentence-transformers is not installed. Run:

pip3 install sentence-transformers --break-system-packages

EMC will fall back to lexical-only recall in the meantime — GRACE still works, just without semantic search.

sqlite-vec not available warning at startup

pip3 install sqlite-vec --break-system-packages

Same fallback as above — lexical recall only until resolved.

GUI can't connect Check that rosbridge is running and that you're using the Jetson's LAN IP, not localhost, when connecting from another device.

Clone this wiki locally