Dice Question Streamline Icon: https://streamlinehq.com

Incorporate object orientation into VLA evaluation metrics for the Move near task

Develop and validate evaluation metrics for Visual Language Action models that explicitly incorporate object orientation in tasks where orientation is critical—such as the Move near task—by first identifying which tasks require orientation assessment and then integrating orientation into the uncertainty and quality metrics so that final orientation correctness is accurately captured.

Information Square Streamline Icon: https://streamlinehq.com

Background

During correlation analysis, the authors found that several metrics failed to capture whether the final orientation of manipulated objects was appropriate in the Move near task, despite orientation being critical for safety and task acceptability in that setting.

They propose leveraging simulator access to object poses and orientations to extend metrics to account for orientation-sensitive tasks, but acknowledge that defining when and how to integrate orientation remains unresolved.

References

Since the simulators usually have access to both the initial and final positions, as well as orientation of the objects, a promising direction for improving our metrics would be to identify tasks where the orientation is important and, subsequently, integrate orientation in our metrics. This remains an open challenge that will be targeted in the future.

Evaluating Uncertainty and Quality of Visual Language Action-enabled Robots (2507.17049 - Valle et al., 22 Jul 2025) in Section 6.2 (RQ2 — Correlation)