Getting SmolVLA to Run on a Low-Cost Arm: Training, Remote Inference, and a First Language-Conditioning Probe

We report how a small vision-language-action model (SmolVLA) was brought, step by step,onto a self-built SO-101 arm that had previously been driven by an ACT policy. The papercovers three things. First, the fine-tuning recipe and the two training runs that produced themain checkpoint used in all 8-position rollouts. Second, the inference problem: the model ranat 0.55 policy requests per second on the arm’s Raspberry Pi 5, about 54 times too slow, andmoving it to a desktop GPU served over the network through LeRobot’s asynchronous inferencebrought the policy-request rate to about 6 Hz while leaving the Pi nearly idle and the armexecuting a 20-action chunk between requests. Third, a small probe of language conditioning,in which four tiny datasets that differ in their instruction (and in the demonstrated trajectory foreach instruction) showed that the policy’s behavior changes with the instruction in the intendeddirection, including for a paraphrase whose exact wording was not in training. The one controlledquantitative comparison we run on the 8-position grid is between the two execution modes onthe same machine, same framework version, and same arm: synchronous inference scored 40/80and RTC scored 41/80, a difference of 1.25 percentage points (Fisher’s exact p = 1.00). Morebroadly, we exercised the same recipe and the same checkpoint across three different LeRobotclones, three client machines, and two physical follower arms during the engineering work inSections 4 and 5, and the pipeline continued to produce functional policies; we flag this asan engineering-reproducibility observation rather than a statistical claim, because the full 8-position protocol was only run under one of those combinations. We do not report a controlledcomparison with ACT; the comparison in the final section mixes model and conditions, and theper-position diagnosis is left to a follow-up paper. This is a single-operator study with smallsamples; it is meant as a reproducible record for others on similar budgets.