Event
Proposes Failure Analysis of VLA in Contact-Intensive Tasks and the FACT Method
1 reports · 1 sources
Carlota Parés-Morlans and other researchers analyzed the causes of failure of vision-language-action (VLA) models in contact-intensive manipulation tasks, proposing two failure modes: precision failure (the policy struggles to produce the small, precise corrective actions required during contact) and force failure (the policy cannot respond to contact forces), and proposed a post-training method, FACT, to improve on them. FACT increases low-noise sampling during post-training to improve fine action prediction, and introduces structured force conditioning, using contact state and force history to guide action generation. The paper reports that in evaluations across five contact-intensive tasks and nearly 2,500 real-world rollouts, FACT achieved an average success rate of 66%, outperforming the previous best baseline of 41%; the work received a CoRL 2026 Spotlight.
Report timeline
- FACT post-training method improves VLA success rate in contact-rich manipulation tasks
The paper investigates why vision-language-action (VLA) models fail in contact-rich manipulation tasks that require precise physical interaction, and proposes two failure modes: precision failures (policies struggle to produce the small, precise corrective actions needed during contact) and force failures (policies fail to respond to contact forces).
arXiv cs.RO ·