Trust-Calibrated VLA Policy Post-Training with Evolving Imagination
Overview
MERL, a framework combining policy optimization with an online-updated simulator, improves VLA models post-training. It balances reliability and efficiency using chunk-level trust and simulator reliability metrics. Achieves 79.7% success on LIBERO-PRO tasks and 80.0% on physical-robot tasks with 50 demonstrations per task. arXiv:2610.16517v1 Latest: MERL enhances VLA models by integrating simulator updates and trust-based replay selection, achieving high success rates on perturbed and physical tasks with limited demonstrations. arXiv:2610.16517v1
Generated from attributed reports · Updated 2 hours ago
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
No current evidence-backed claims. Missing information remains not reported.
Evidence dependency changes (1)
- 2026-10-15T19:23:22.185Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsActive Generative Simulation: Meta-Learning What to Simulate for Sim-and-Real Policy Co-Training
The paper presents MetaGenSim, a meta-learning framework that trains an auxiliary rater network to optimize a simulator's data-generation distribution for improved real-world performance.
- arXiv Robotics — research abstractsTrust-Calibrated VLA Policy Post-Training with Evolving Imagination
This research proposes MERL, a framework that couples policy optimization with an online-updated simulator to refine VLA models post-training. The method balances reliability and efficiency by controlling the ratio of imagined data and rollout horizon based on simulator reliability. It also uses chunk-level trust to guide replay selection and optimization. MERL achieves 79.7% average success on perturbed LIBERO-PRO tasks and 80.0% on physical-robot tasks with 50 demonstrations per task.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).