Bayesian Geometric-based Interactions Learning Model for Self-aware Autonomous Agents
Autonomous agents can perceive and act in the environment only by interacting with neighboring agents. This paper proposes a novel probabilistic interaction model that allows the agent to learn the dynamics of the interactive experiences incrementally. The learned model represents the dynamic intera...
Uloženo v:
| Vydáno v: | Signal processing Ročník 239; s. 110237 |
|---|---|
| Hlavní autoři: | , , , , |
| Médium: | Journal Article |
| Jazyk: | angličtina |
| Vydáno: |
Elsevier B.V
01.02.2026
|
| Témata: | |
| ISSN: | 0165-1684 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | Autonomous agents can perceive and act in the environment only by interacting with neighboring agents. This paper proposes a novel probabilistic interaction model that allows the agent to learn the dynamics of the interactive experiences incrementally. The learned model represents the dynamic interaction variables within a joint Generalized State (GS) space. This allows to represent the contextual information of the agent’s experiences in the form of time-varying probabilistic graphical models whose variables are geometrically interpreted as attractors . Such a generative model is defined as Interactive Hierarchical Generalized Dynamic Bayesian Network (IH-GDBN). Various interaction models describing the agents’ experiences are collectively stored as bio-inspired memory layers of the agent’s Autobiographical Memory (AM). Knowledge stored inside AM is accessed by the Bayesian inference method called Interactive Geometrical Markov Jump Particle Filter (IG-MJPF) and is able to make inferences of future interaction states based on learned IH-GDBNs. Moreover, such filters are enriched to detect anomalies as effects of unknown geometrical forces, i.e., deviating from the predictions of a model called as Generalized Errors (GEs). The estimation of GEs allows the agent to learn the new models incrementally by evolving the respective layers of AM to adapt the changing interaction situations. The proposed method is tested in real-time, complex overtaking experiments involving two Autonomous Vehicles (AVs). Future work will extend these experiments to scenarios with more than two vehicles to better reflect multi-agent traffic dynamics. Two different sensory modalities are employed to show, how the AM memory layers can be learned from the exteroceptive, i.e., positional trajectories (called odometry module) and proprioceptive sensors, i.e. steering angle and rotors’ velocity. Performance of the proposed method highlights the detection capabilities as well as the ability to learn explainable incremental successive models within the AM. Codes related to this work can be accessed via https://github.com/Hafsa-Iqbal/Interaction-Modeling.
•Geometric reasoning-based unsupervised data-driven method to infer interactions between AV states.•Geometric-based approach to estimate active/passive Interactive Transition Matrix for AVs.•Validation of method with complex real-world AV overtaking experiments.•Estimation of abstraction-level anomalies to identify novel interactions.•Self-aware interaction model with evolving memory for continual knowledge integration. |
|---|---|
| ISSN: | 0165-1684 |
| DOI: | 10.1016/j.sigpro.2025.110237 |