Act-Now: Autonomous Cognitive Technologies for Novelty in Open Worlds
Abstract
The main objective of the project was to develop a versatile novelty-aware cognitive robotic architecture that will enable artificial agents to cope with "unknown unknowns". Specifically, we aimed to develop inference-based and learning-based methods for novelty detection, combining statistically trained sub symbolic models with knowledge-based symbolic models and utilizing a mixture of symbolic and statistical inference for detecting deviations from expectations. We used three TA1 tasks to develop these algorithms: Polycraft, Monopoly, and NLP (the last was removed in the final phase of the project). Another main goal was to also characterize the detected novelties and determine how to cope with them based on the effects they had on the agent's task. We also planned to integrate all algorithms for novelty detection, characterization, and accommodation into the unified architecture framework to be able to utilize them in future work in different domains.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 11, 2023
- Accession Number
- AD1225550
Entities
People
- Chitta Baral
- Jivko Sinapov
- Liping Liu
- Matthias J Scheutz
- Michael C. Hughes
- Subbarao Kambhampati
Organizations
- Tufts University