AutoACAP

Autonomous AI-driven optimization of ACAP applications on Axis camera hardware, inspired by Karpathy's autoresearch.

An AI agent rewrites person detection ACAPs in C and Rust, deploys each variant to a real AXIS Q6358-LE camera (ARTPEC-9), benchmarks head-to-head, and ratchets toward the fastest implementation — overnight, unattended.

How It Works

The repo has three layers:

• program.md — Instructions for the AI agent. The human writes this.
• src/c/app.c and src/rust/src/main.rs — The agent modifies these.
• evaluate.sh — Fixed benchmark harness. Builds, deploys, benchmarks, scores. Never modified.

Each experiment:

1. Agent proposes a change to app.c or main.rs
2. Git commits the change
3. evaluate.sh builds, deploys to camera, runs 60s benchmark
4. If FPS improved (and mAP >= 0.40) → keep commit
5. If FPS same or worse → git reset
6. Log to results.tsv, repeat forever

Quick Start

# 1. Verify camera access
ssh root@192.168.1.33

# 2. Pull ACAP Native SDK Docker image
docker pull axisecp/acap-native-sdk:12.9.0-aarch64-ubuntu24.04

# 3. Validate model on camera DLPU
# (copy model to camera and test with larod)

# 4. Capture benchmark ground truth
# (see benchmark/scripts/capture_benchmark.sh)

# 5. Run a manual experiment
./evaluate.sh c 192.168.1.33 <password>

# 6. Start autonomous research
# Point Claude Code at this repo:
# "Read program.md and let's kick off a new experiment!"

Project Structure

autoacap/
├── program.md              # Agent instructions (human writes)
├── evaluate.sh             # Fixed benchmark harness (DO NOT MODIFY)
├── results.tsv             # Experiment log (untracked)
├── src/
│   ├── c/                  # C variant
│   └── rust/               # Rust variant
├── benchmark/              # Video, ground truth, scripts
├── models/                 # .tflite model files
├── deploy/                 # Camera deploy/start/stop scripts
└── analysis/               # Post-run Jupyter notebook

Target Hardware

• Camera: AXIS Q6358-LE (ARTPEC-9, aarch64, AXIS OS 12)
• IP: 192.168.1.33
• Model: SSD MobileNet V2 INT8 (.tflite, DLPU Chip ID 12)
• Stream: 640x480, NV12, 15 FPS

Metric

• Optimize: FPS (higher is better)
• Gate: mAP >= 0.40 (discard if below)
• Track: latency_p95, memory_mb, cpu_pct, binary_kb

License

MIT