A real time motion first autonomy runtime built to treat latency as a law, not a suggestion.
AEROS is an authoritative autonomy runtime for constrained edge hardware. It utilizes an asynchronous heartbeat loop to decouple low level control from high frequency telemetry. The architecture implements three layers of vehicle authority: PID based stabilization, temporal state inference through grayscale delta encoding, and geometric safety through conformal algebra.
It exists to answer one question: given a single camera and constrained edge hardware, what is the minimum viable perception pipeline required to maintain vehicle stability?
| Metric | Baseline Performance |
|---|---|
| End to End Latency | 11.4 ms (Verified Field Trial 001) |
| Perception Cadence | 30 FPS Fixed Authoritative Heartbeat |
| Target Hardware | iPhone Sensor Node |
| HUD Performance | 60 FPS Zero State Render Loop |
AEROS prioritizes motion over color. Its purpose is to perceive geometry through change, maintain temporal state, and predict near future motion on edge devices where compute is a limited resource. The system is organized into three specific layers of execution:
- Muscle: PID based stabilization and low level control authority.
- Intuition: A lightweight streaming state space model (SSM-lite) that reasons over temporal contrast instead of isolated RGB textures.
- Shield: A Conformal Geometric Algebra safety layer that evaluates vehicle position inside a valid motion tube.
The runtime uses an authoritative ownership model for each active video source. This central loop owns frame acquisition, temporal state, predictor state, control output, and telemetry generation.
Unlike standard pipelines, AEROS decouples logic from I/O. WebSocket subscribers are passive read only consumers. They are never allowed to apply control or own authoritative state. If the client connection drops or the network signal degrades, the physics and perception loop continues at its fixed rate. Camera hot swapping allows the engine to switch sensors on the fly without killing the heartbeat or severing the telemetry connection.
The frontend is a high fidelity flight instrument built for hardware efficiency. It bypasses React state entirely to hit high frequency motion targets under a strict 5ms budget.
- Aviation Window Tapes: Digital readouts sit in fixed glass boxes while scalar tracks physically slide behind them to replicate mechanical jet avionics.
- Kinetic Bore Sight: A functional indicator featuring a Velocity Vector Ghost Icon that shifts laterally based on raw heading rate to show actual physical drift.
- Zero State Animation: All sliding tapes and pulsing brackets update via direct DOM mutations inside a
requestAnimationFrameloop. I usecreateImageBitmapanduseRefto eliminate garbage collection pressure and prevent the main thread from choking.
I have transitioned AEROS from simulation to real world hardware. Trial 001 utilized an iPhone sensor node as a surrogate drone to evaluate the authoritative runtime in an outdoor administrative environment.
Performance Success:
- Latency Law Verified: The system achieved a median end to end latency of 11.4ms, confirming that the 5ms stage budget is sustainable on physical hardware.
- Heartbeat Authority: The backend loop maintained absolute synchronization. Sequence IDs remained continuous and the runtime loop stayed active even when the network signal degraded at range.
Engineering Gaps Identified:
- Transport Limitations: Standard WiFi proved insufficient at range. FPS dropped significantly at approximately 80 steps, identifying the need for dedicated low level radio transport.
- Perception Jitter: The legacy RGB CNN showed heading instability while stagnant. Hand vibrations and surface reflections created false motion signals, confirming that grayscale temporal encoding is the only stable path for edge autonomy.
I value technical honesty. This system does not attempt SLAM, path planning, or ROS2 integration. The model is currently trained on synthetic corridor data.
macOS Native Simulation is Disabled: Upstream Bullet physics currently fails to build against recent Apple SDKs. If you are running AEROS natively on macOS, simulation will fail. The camera feed, UI, and inference engine work perfectly, but physics simulation requires a Docker Linux environment.
I am actively transitioning to the AEROS.vOmega architecture.
- Temporal Eyes: Computing temporal delta encodings from grayscale frames to drop reliance on static RGB textures.
- Predictive Brain: Replacing the frame wise CNN with a lightweight State Space Model (SSM-lite) that maintains a hidden state.
- Algebraic Shield: Prototyping a Conformal Geometric Algebra (CGA) safety layer to provide mathematical certainty over branchy heuristics.
- Sim to Real Bridge: Deploying ONNX INT8 quantized models to Pi 5 hardware and talking to flight controllers via MAVLink SITL.
# Terminal 1: The Heartbeat
uvicorn api.main:app --host 0.0.0.0 --port 8000 --reload
# Terminal 2: The Instrument
cd web && npm startFor headless simulation (bypassing macOS local build failures):
docker compose up --build apiThe flight instrument mounts at http://localhost:3000.
This documentation reflects AEROS as it exists today in the main branch.