Bilevel Policies Combine Symbolic Planning with Imitation Learning for Long-Horizon Tasks
A new arXiv preprint (2605.15975) proposes a bilevel policy framework for embodied AI agents to solve long-horizon planning problems. The approach combines a high-level symbolic policy for efficient, interpretable planning with a low-level neural policy learned from demonstrations for fine motor control. The high-level policy operates over symbolic world models, while the low-level policy handles continuous manipulation. This hybrid method aims to overcome the limitations of pure imitation learning in generating extended plans, leveraging the strengths of both abstraction levels. The paper does not specify experimental results or benchmarks.
Key facts
- arXiv preprint 2605.15975 proposes bilevel policies for long-horizon planning
- Combines high-level symbolic policy with low-level neural imitation learning policy
- High-level policy enables efficient and interpretable long-horizon planning
- Low-level policy handles fine motor control and manipulation in continuous environments
- Aims to overcome limitations of imitation learning alone for long-horizon tasks
- No experimental results or benchmarks are reported in the abstract
Entities
Institutions
- arXiv