Doob-Based Generative Policy Optimization
We propose Doob Policy Optimization (DoobPO), a framework that leverages Doob's $h$-transform for likelihood-free reinforcement learning fine-tuning of generative policies.
ProofPaper ↗
Key points
- Diffusion and flow models provide expressive policy classes for robotic and sequential decision-making systems.
- DoobPO separates policy improvement from generative policy learning.
- We prove that the practical conditional drift matching objective is gradient-equivalent to the ideal Doob drift matching objective.
- Across MuJoCo continuous-control benchmarks, DoobPO achieves strong performance against competitive diffusion- and flow-policy baselines, showing a principled and practical route to RL fine-tuning for generative policies.
Sources (1)
- [1]Doob-Based Generative Policy OptimizationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 11:31 PM
We propose Doob Policy Optimization (DoobPO), a framework that leverages Doob's $h$-transform for likelihood-free reinforcement learning fine-tuning of generative policies.
Diffusion and flow models provide expressive policy classes for robotic and sequential decision-making systems.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026RL-IF: Tracking Non-local Training Data Influence for Online Reinforcement Learning
- Oct 20, 2026Extreme Low-State Quantized Transformer for Electricity-Load Forecasting
- Oct 20, 2026Fine-tuning GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod
- Oct 20, 2026Context-Aware Prompt Injection Detection via Small Language Models in RAG Pipelines: nokast-secureRAG
- Oct 20, 2026Reframing Object-Centric Reinforcement Learning: From Absolute to Relative Representations
- Oct 20, 2026Training the Whole Agent, Not Just the Model