Robust Task-Aware State Estimation from Pre-trained Perception
Summary
- Researchers propose Robust Task-Aware State Estimation from Pre-trained Perception to address the uncertainty of Vision-Language Models in agent state estimation.
- The method constructs a confidence-based Minimal Belief Set using underlying perception model probabilities and task constraints.
- The proposed Probability-Ranked Valid State Search algorithm builds this belief set with polynomial complexity and correctness guarantees.
Why it matters
This approach bridges the gap between brittle single-guess perception and intractable full probability distributions, improving downstream agent decision-making using pre-trained models.