THE MECHANISM

Baseline → bake-off → SmolVLA → async serve

03Baseline first
→
04Smoke-test before marathon training
→
05Async serving needs a buffer invariant
FAILURE → REPAIR
Common trap

“Lower training loss selects the shipping model.”

Better model

Shipping depends on real success, recovery, latency, and operational stability—not only the optimization curve.

YOUR NEXT GATE

Train with gates

  1. 1Run lerobot-train --help and save the installed version.
  2. 2Train ACT on the validated dataset.
  3. 3Evaluate ten trials per object position.