Sound POMDP Synthesis with LTL Objectives via Reward Shaping
A research paper on arXiv (2605.12581) introduces a novel reward-shaping mechanism for synthesizing autonomous agents in partially observable Markov decision processes (POMDPs) with Linear Temporal Logic (LTL) objectives. The approach dynamically generates belief-dependent rewards grounded in certified LTL satisfaction, integrated into an enhanced Monte Carlo Planning framework. Experiments show it thrives where existing methods fail, addressing the undecidability of qualitative LTL verification in POMDPs.
Key facts
- arXiv paper 2605.12581
- Title: Ensuring Logic in the Fog: Sound POMDP Synthesis with LTL Objectives
- Addresses synthesis of autonomous agents under uncertainty with LTL constraints
- Proposes a sound reward-shaping mechanism for POMDPs
- Rewards are belief-dependent and grounded in certified LTL satisfaction
- Integrated into an enhanced Monte Carlo Planning framework
- Experiments demonstrate success in scenarios where existing methods fail
- Bridges gap between LTL specification and quantitative synthesis in POMDPs
Entities
Institutions
- arXiv