LeRobot deployment adapter

LeRobot trains the policy, runs cameras, and talks to the robot. FlowEdge is the native action head behind that plugin: converted diffusion_pusht first, then a cached SmolVLA expert. The problem this solves is the last milliseconds — a fixed U-Net, a period, and an explicit miss policy — without putting e-stop or joint limits inside Core.

Boundary

LeRobot	FlowEdge
Cameras, RGB/state encoder, normalization, history	Fixed-memory U-Net / expert
Robot driver, joint limits, e-stop	--on-miss hold|drop|raise
Training	Converted .safetensors

Action slice: actions[observation_steps - 1 : observation_steps - 1 + action_steps]. input_mode="visual" uses the source encoder. A flattened condition remains valid. Policy evaluation.

Install

python -m pip install -e integrations/lerobot
python -m unittest discover -s integrations/lerobot/tests

from flowedge_lerobot import FlowEdgeDiffusionPolicy
policy = FlowEdgeDiffusionPolicy.from_checkpoint(
    "models/diffusion_pusht.flowedge.safetensors", threads=4
)

Period loop (sim, Jetson, SO-100)

python -m flowedge_dev pipeline rollout models/diffusion_pusht.flowedge.safetensors \
  --steps 20 --threads 4 --period-ms 10 --on-miss hold

CUDA Core (FLOWEDGE_BACKEND=cuda), same miss contract:

python -m flowedge_dev pipeline rollout models/diffusion_pusht.flowedge.safetensors \
  --steps 20 --threads 1 --period-ms 10 --on-miss hold --device cuda

--device cuda requires a CUDA Core binary and Engine.cuda_resident. A 4 GB card that cannot upload the U-Net stays on CPU kernels; that is a load success, not a CUDA replay. FLOWEDGE_CUDA_REQUIRED=1 fails closed. GTX 1650, 10 ms hold: 20 / 20 misses. Not a Jetson/ARM claim.

--on-miss	On overrun
hold	Repeat last sent action; zeros before the first on-time send
drop	Skip send_action
raise	DeadlineMissed after stop

Jetson: scripts/edge_dp_rollout.sh. Attach JSON to issue #70. SO-100: implement reset / observe / send_action / stop and pass encode_condition into run_rollout. CI uses a fake robot. ARM64 CI is architecture validation, not a board claim.

--policy.type=flowedge is DP. --policy.type=flowedge_smolvla is the native expert plus a LeRobot VLM cache — not a native VLM. Tenstorrent is a separate backend, not this plugin.