New Embedding Temporal Logic lets AI monitor perception in real time
Researchers propose ETL to monitor autonomous systems directly in learned embedding spaces
Deep Dive
A new temporal logic called Embedding Temporal Logic (ETL) monitors perception-based autonomous systems directly in learned embedding spaces. ETL defines predicates through distances between observed and target embeddings, enabling high-level perceptual concepts like visual goal similarity or semantic region avoidance. The approach includes a conformal calibration procedure for reliable predicate evaluation, validated across multiple manipulation environments with strong agreement to ground-truth semantics.
Key Points
- ETL defines temporal logic predicates using distances between observed and target embeddings, bypassing brittle low-dimensional mappings.
- Conformal calibration provides safety-oriented predicate evaluation with statistical reliability bounds.
- Tested across multiple robotic manipulation environments with strong agreement to ground-truth semantics for temporally composed behaviors.
Why It Matters
ETL enables safer autonomous systems by monitoring perceptual decisions directly in embedding spaces, reducing failures from semantic misalignment.