Learning Bilevel Policies over Symbolic World Models for Long-Horizon Planning

Researchers introduced BISON, a bilevel planning system for embodied AI that combines low-level imitation learning with high-level symbolic world models.

The system learns manipulation skills from demonstrations, then uses symbolic abstractions to plan over long horizons more effectively.

On extended MetaWorld benchmarks, BISON generalizes better and is more efficient than competing end-to-end approaches, especially on harder multi-object tasks.