Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them

2026-08-02

AI Summary

Existing Vision-Language-Action (VLA) models exhibit poor performance on contact-rich manipulation tasks, yet the underlying causes of their failures remain unclear. This work systematically identifies and disentangles two distinct failure modes: precision failures stemming from mismatches in flow-matching training strategies, and force failures arising from the unique structural properties of force signals. To address these issues, we propose FACT, a novel approach that integrates tailored flow-matching optimization with explicit modeling of force signal structure, complemented by a force-augmented architecture and training regularizations. Evaluated across five real-world contact-intensive tasks through nearly 2,500 trials, FACT achieves an average success rate of 66%, substantially outperforming the previous best baseline at 41%, thereby significantly enhancing the reliability and success of physical interaction.

Abstract

We address the problem of understanding when and why Vision-Language-Action models struggle with contact-rich manipulation tasks that require precise physical interaction. Prior work has primarily focused on addressing contact failures through force-augmented architectures and training-time regularizers, yet the root causes of these failures remain underexplored. We identify two distinct failure modes underlying this gap. Precision failures are rooted in a flow-matching policy training mismatch, and force failures arise from the distinctive structure of force signals. We address each failure mode with a targeted mechanism and combine them into FACT, which achieves 66% average success rate across five contact-rich tasks against 41% for the best prior baseline, in an evaluation spanning almost 2,500 real-world rollouts.

Problem

• Vision-Language-Action models
• contact-rich tasks
• manipulation failures
• precision interaction
• force signals

Innovation

• Vision-Language-Action models
• contact-rich manipulation
• failure mode analysis
• force signal structure
• flow-matching policy