Plug in your robot. Type what you want. AETHER finds every sensor and motor, writes the code to drive them, and the FDIR layer keeps it safe.
AETHER is a prompt-driven robotics framework. The discovery layer auto-detects every connected device · cameras, GPIO motors, flight controllers, I²C sensors. The Claude API turns your prompt into a working program against the real hardware manifest.
The PPO-based FDIR layer is the safety net: it watches every loop in real time and recovers from faults the moment they appear · drift, stuck motors, sensor failure. The net learns, the rules backstop.
ToolDiscovery probes for cameras, GPIO pins, flight controllers, I²C sensors, network interfaces and AI models. No config files required.
ToolBuilder constructs only the tools that will actually work on the present hardware. Everything downstream is gated on this manifest.
NavigationEngine picks the right autonomy level; the planner avoids ever planning with unavailable hardware. Hierarchical task expansion.
Execution is wrapped in continuous fault-detection-identification-recovery. PPO detects, temporal validator confirms, planner issues recovery.
15-dimensional observation in → 64 → 64 hidden → 8 fault classes out. Trained on real telemetry, not idealised simulation noise. Outperforms threshold rules under genuine sensor drift.
$ aether --status ● PPO detector nominal 15→64→64→8 ● Temporal validator nominal w=8 frames ● Safety backup armed 5 rules ● Recovery planner idle last: 0
Tell AETHER what your robot should do. It probes USB, I²C, GPIO, MAVLink, and on-device models, builds a capability manifest, then asks the Claude API to write code against exactly what's connected.
manifest@boot + webcam:usb-0 (opencv) + motors:gpio (rpi.gpio) + fc:mavlink (/dev/ttyACM0) + imu:i2c-0x68 (mpu6050) + ai:yolov8n (local)
Detection, validation, planning, execution, safety and recovery agents share state over a deterministic message bus.
DET ─► VAL ─► PLN │ │ │ ▼ ▼ ▼ SAF ◄─ EXE ◄─ REC
A rolling 8-frame window discards transient PPO spikes. Recovery only triggers on durable, validated detections · false positives stay low.
window t-7 ... t class: 0 0 3 3 3 0 3 3 → ▲ MOTOR class: 0 5 0 0 0 0 0 0 → ─ idle class: 3 3 3 3 3 3 3 3 → ▲ MOTOR
Verified across thousands of operational steps on quadrotor, ground rover and arm. Real hardware outperforms simulation · genuine noise teaches the net.
Quadrotor 2,140 steps DR 100% Rover 3,802 steps DR 100% Arm 1,160 steps DR 100% Sim baseline DR 94%
Six cooperating agents read off the shared state bus. The DRL detector is the only ML-trained component · everything else is deterministic, auditable, and fails closed. Latency from sensor frame to recovery command is bounded.
Real hardware outperforms simulation · the physical system encounters genuine sensor noise that the PPO net learns to distinguish from actual faults.
Wheels for Python 3.10–3.14, Linux / macOS / Raspberry Pi OS. Calibration writes a hardware manifest to ~/.aether/manifest.json. Agent mode runs the full FDIR loop.
Aether is provider-agnostic. Drop in an API key, the framework discovers your sensors and motors, and the model writes the code against the live capability manifest. Claude is the default; OpenAI, Gemini, DeepSeek, and Kimi all plug in through the same provider block.
The default integration. Aether ships pre-wired for Anthropic's Claude. Best results for tool-grounded code generation, multi-step planning, and JSON-mode calls against the discovered capability manifest. Sonnet is the recommended model for most rigs.
# ~/.aether/config.toml provider = "anthropic" model = "claude-sonnet-4-5" api_key = "$ANTHROPIC_API_KEY" max_tokens = 8192
Full support through the OpenAI SDK. Function calling maps directly onto Aether's capability tools. Works with the latest GPT family and the 4o tier for cheaper local development.
Gemini 2.5 Pro and Flash both supported. The long context window comfortably fits hardware manifests with dozens of sensors. Auto-fallback is configurable from the provider block.
Strong code-gen benchmarks at a fraction of the cost. Reasoning output is parsed for tool-call intents and routed back through the same scheduler. Useful for offline lab rigs.
Kimi K2 sits behind an experimental flag. Excellent for very long planning horizons. A small shim translates between the Aether scheduler and the Moonshot HTTP API.
Export it as an environment variable or write it into the Aether config. Every provider reads from the same block, so switching models is one line.
$ export ANTHROPIC_API_KEY=...
Run calibration once. Aether probes USB, I²C, GPIO, MAVLink, and on-device models, then writes a capability manifest that grounds every prompt.
$ aether --calibrate
The prompt, the manifest, and the FDIR constraints are sent together. The model returns code. FDIR wraps that code on execution and keeps it safe.
$ aether do "patrol the room"
A robotics engineer building the autonomy layer he wished existed.
UIUC