Proposal: out-of-tree fixed-shape FlowEdge deployment plugin

Proposal

FlowEdge now provides an optional, out-of-tree lerobot_policy_flowedge deployment plugin for a converted fixed-shape LeRobot Diffusion Policy. It keeps training, observation encoding, robot drivers, limits, and emergency stop in LeRobot or the application; FlowEdge owns only the fixed-memory action-policy path.

Runnable evidence

• Plugin and bounded simulator rollout: https://github.com/elprofesoriqo/FlowEdge/tree/main/integrations/lerobot
• Native model-path/ONNX decision table: https://github.com/elprofesoriqo/FlowEdge/blob/main/docs/guides/model-import.md
• Cross-platform CI, ARM64/NEON build/test, and deterministic PyTorch-to-ONNX Runtime parity: https://github.com/elprofesoriqo/FlowEdge/pull/133

The simulator command reports startup time, p50/p99/max timing, missed deadlines, high-water RSS, platform, and machine. The native runtime preserves zero post-setup hot-path allocations; optional ONNX Runtime stays outside that contract.

Deliberate limits

This is not a request to add a graph runtime, training stack, robot driver, or ACT/VLA implementation to LeRobot or FlowEdge. The initial supported path is converted diffusion_pusht with a flattened preprocessed condition. A real SO-100/SO-101 pilot is still pending.

Feedback requested

1. Is an out-of-tree lerobot_policy_* package the right deployment integration boundary?
2. Which processor/artifact contract should a deployment adapter treat as stable?
3. Which simulator or SO-100/SO-101 fixture would be most useful for a first reproducible pilot?

I will maintain the companion package and keep any hardware-specific code out of FlowEdge Core/Relay.