- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This work addresses a core problem, the vulnerability of VLA policies when crucial spatial relations are occluded or missing in current views. This can cause both nominal actions and tentative inference-time corrections to carry high uncertainty. To solve this problem, the authors introduce "Ask the World", a correction-guided active perception framework that treats a tentative correction as a decision-directed visual query to identify key spatial relations that must be resolved. The framework uses LingBot as the frozen base VLA policy to gather nominal action chunks, paired with Astra as an inference-time visual reasoning assistant to process local corrections and manages active view acquisition.
Critical strengths
(a) Novel formulation of action corrections as visual queries. In this work, instead of treating inference-time corrections isolated as direct trajectory modifications or relying on generic visual exploration, authors innovatively treat a tentative correction as a decision-directed visual query. This novel concept explicitly links active perception to the spatial relation needed to accept or reject an intervention.
(b) Inference-time assistance without retraining base models. The proposed framework operates around a frozen VLA policy and uses a general-purpose visual reasoning assistant without updating the parameters of model and no need retraining the base policy. This provides a flexible, modular pathway to enhance policy robustness at inference stage.
Critical weaknesses
(1) My first concern is that: the evaluation relies only on a virtual camera in the RoboTwin simulator to acquire new viewpoints. The authors also acknowledge that this bypasses the real-world physical constraints, collision risks, trajectory execution time, and hardware costs associated with physically moving an embodied camera.
(2) Despite performance gains are significant on spatially occluded tasks, the macro-average improvement across all 16 tasks over the base frozen LingBot policy is relatively margin (93.13% vs. 91.38%). On several tasks (e.g., Grab Roller, Scan Object), the system actually performs even slightly worse than the baseline policy.
Questions for authors
Q1: How does the framework balance the physical time, energy, and collision risks of moving a physical camera when transitioning from simulator virtual views to real world scenarios?
Q2: Why authors designed Astra to introduce unnecessary corrections that drop performance on tasks where the base policy already perform well (e.g., Grab Roller falling from 100% to 95%)?