Overview

FlowEdge is a small stack with one job per layer. The point is worst-case latency on a robot period, not average throughput.

A general graph runtime would accept more architectures with less hand work. It would also reintroduce allocation, dispatch, and a working set that is hard to bound — exactly the variance a control loop cannot absorb. So the model is hand-assembled: convert tensors into known names, carve one arena, call kernels through one header. ADR 0005.

Hardware runs the loop, FlowEdge runs the policy, LeRobot trains and drives the robot Whole FlowEdge flow from checkpoint to motors Optional Relay flow from sensor or VLA through shared-memory rings to Core Layers from plugin down to the arena

Two product heads share that stack:

  • Flow matching — a short ODE (dx/dt = v(x,t \mid c)) from noise to action. Mamba (or a caller-supplied encoder) produces (c).

  • Diffusion Policy — DDIM on a Conv1D U-Net. LeRobot’s RGB encoder stays outside Core.

Transformer conversion and a cached SmolVLA expert are incubators, not those heads.

Mamba plus flow matching, or LeRobot encoder plus DP U-Net tokens or condition, backbone, flow ODE, action chunk

  • Arena. One bump slab. Load may allocate it; the hot path does not.

  • Loader. Zero-dep .safetensors mmap. F32 and BF16.

  • Runtime. engine_runtime owns persistent state, model assembly, SPMC pool. C ABI is a thin adapter.

  • Kernels. One header. CPU today. FLOWEDGE_BACKEND=cuda keeps Mamba, flow, and Diffusion Policy heads on device. Vulkan / Tenstorrent (Metal) stay planned.

  • Backbone. Prefix → condition. Mamba today (constant-size state). Transformer is a CPU decoder fixture, not a policy.

  • Head. Condition + noise → action. Flow matching or Diffusion Policy. An external encoder can skip the backbone.

Split the ODE with flow_begin / flow_advance if the robot has to yield. Cooperative execution. Model code calls matmul and discretize_and_scan, never a device API. Footprint is computed at load.