FOUNDER: Grounding Foundation Models in World Models for Open-Ended Embodied Decision Making
TLDR
FOUNDER integrates foundation models with world models for open-ended embodied decision making, using a mapping function and temporal distance reward.
Reasoning
The paper presents a novel framework combining foundation models and world models, with strong empirical results on multi-task benchmarks. However, the abstract lacks details on limitations and scalability, and the reliance on offline data may limit real-world applicability.
Read-first score
Read-first score 53.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.
Field roles
Rank sensitivity
Stability: volatile; rank range: 461.
Keyword Scores
Deep Analysis
Innovations
- Integration of Foundation Models and World Models for open-ended embodied decision making in a reward-free manner
- Learning a mapping function that grounds FM representations into the WM state space, enabling inference of physical states from external observations
- Using predicted temporal distance to the goal state as an informative reward signal for goal-conditioned policy learning through imagination
Methodology
FOUNDER learns a mapping function that aligns Foundation Model representations with the state space of a World Model, allowing the agent to infer its physical state from observations. This mapping enables goal-conditioned policy learning via imagination, where the predicted temporal distance to the goal state serves as a reward signal, eliminating the need for external rewards.
Key Results
FOUNDER achieves superior performance on multi-task offline visual control benchmarks, particularly in tasks with complex observations or domain gaps where prior methods struggle. The learned reward function is empirically validated to be consistent with ground-truth rewards.