FEP-Diff: Free Energy Principle for Agent-Centric Trajectory Prediction
FEP-Diff, a novel framework, introduces agent-focused trajectory prediction based on the Free Energy Principle, tackling the shortcomings of current methods that depend on global state assumptions and lack cognitive behavioral limitations. This method employs a dual-branch spatiotemporal encoder to derive ego-motion dynamics and social interaction signals from local observations. Additionally, a goal-conditioned belief learner deduces multimodal latent belief distributions, optimized through a free-energy objective with a social consistency constraint within the local environment. The objective of this research is to produce cognitively realistic predictions under practical constraints, enhancing the feasibility of deployment and physical realism in actual scenarios. The paper can be found on arXiv with the identifier 2605.25748.
Key facts
- FEP-Diff is a trajectory prediction framework grounded in the Free Energy Principle.
- It uses a dual-branch spatiotemporal encoder for ego-motion and social cues.
- A goal-conditioned belief learner infers multimodal latent belief distributions.
- The objective is optimized via a free-energy objective with social consistency constraint.
- Existing methods rely on global state assumptions and lack cognitive constraints.
- The approach aims for cognitively plausible predictions under partial observability.
- The paper is published on arXiv with identifier 2605.25748.
- The work addresses deployment feasibility and physical plausibility.
Entities
Institutions
- arXiv