What Matters in Quality-Aware VLA Pretraining: A Benchmark and Empirical Study
arXiv cs.AI · 14 Oct 2026, 03:59 AEST
A study investigates how demonstration quality impacts vision-language-action (VLA) pretraining to enhance performance in target-task fine-tuning. The research introduces AXIS-Bench for controlled comparisons and finds significant improvements in success rates through various conditioning and filtering methods.
Key facts
- We compare weighting, filtering, and conditioning across reward definitions and label granularities using 5,000 naturally suboptimal human demonstrations and 24 target tasks.
- Across three VLAs, we find that approach-aware episode conditioning improves mean success over plain behavioral cloning.
- In our real-world experiment, simulation pretraining raises success from 9.6% to 31.7%.
Read aloud
A study investigates how demonstration quality impacts vision-language-action VLA pretraining to enhance performance in target-task fine-tuning. The research introduces AXIS-Bench for controlled comparisons and finds significant improvements in success rates through various conditioning and filtering methods. We compare weighting, filtering, and conditioning across reward definitions and label granularities using five thousand naturally suboptimal human demonstrations and twenty four target tasks. Across three VLAs, we find that approach-aware episode conditioning improves mean success over plain behavioral cloning. In our real-world experiment, simulation pretraining raises success from nine point six percent to thirty one point seven percent.
Sources in this cluster
- What Matters in Quality-Aware VLA Pretraining: A Benchmark and Empirical Study
arXiv cs.AI · 14 Oct 2026, 03:59 AEST