Two papers target VLA generalization and wording sensitivity

Two Oct. 7 papers improve vision-language-action policies: DRIVE raises out-of-domain success via diversity rewards, while instruction rephrasing lifts frozen models without retraining.

Two arXiv papers posted on 7 October 2026 examine generalization of vision-language-action policies, including pi0 and pi0.5, but they propose different fixes rather than conflicting measurements of one method. DRIVE converts diversity among successful behaviors into an intrinsic reward during reinforcement-learning fine-tuning; versus vanilla RL fine-tuning, average out-of-domain performance rises by 5.3 points on pi0 and 2.0 points on pi0.5 across LIBERO-Plus, ManiSkill3, and RoboTwin 2.0. On a dual-arm AgileX PiPER-X platform, DRIVE raises average out-of-domain success from 64.1% to 73.3%. A separate study finds sharp language sensitivity: pi0.5 succeeds 100% of the time for "switch on the stove" but only 2% for "switch on the hot plate," and a rephrase-augmented pi0 checkpoint still swings by up to 61 points. An oracle phrase search nearly closes a 21-point in-distribution versus out-of-distribution gap; distilling 10 to 20 rephrasing rules and rewriting each instruction once improves frozen pi0 by 16–27% relative on twelve held-out tasks and lifts pi0.5 in-finetune LIBERO success from 93.6% to 97.8% without retraining. The reports do not disagree on shared figures; they describe distinct interventions.