Product Direction

FlowEdge is the boundary between a trained policy and robot control: convert once, fixed RSS, a period, a replayable action. Hardware runs the loop. FlowEdge runs the native head. LeRobot (or your stack) trains, encodes observations, and owns e-stop.

Hardware, FlowEdge, LeRobot

Keep

Leave

Hardware: CPU now; Jetson/ARM via plugin; CUDA Mamba, flow, and DP heads; matched GPU replay (#163) and CUDA Core period loop (#164) on GTX 1650

Robot e-stop and joint limits

FlowEdge: flow matching, DP U-Net, Transformer decoder fixture, optional Relay

Training, datasets, graph compilers

LeRobot: plugin, encoders, drivers

Native kernels and the arena

The sequence is convert → matched CPU replay vs PyTorch → Transformer fixture (#10) → CUDA device-resident flow (#161) and DP (#162) → matched GPU replay (#163) → CUDA DP conv occupancy (#174, #176, #179) → a period loop on Jetson or SO-100 with miss counts. Tenstorrent stays behind that GPU evidence. Success is an external team doing that without FlowEdge training code.

convert, CPU replay, Jetson or SO-100, same contracts