OGPO: Offline Goal-conditioned Policy Optimization for Recoverable Vision-Language-Action Models
Vision-language-action (VLA) models have shown strong promise in robotic manipulation, yet their post-training still predominantly relies on trajectory-level supervised imitation over robot demonstrations.
This paper introduces OGPO, an offline goal-conditioned policy optimization framework designed to improve the robustness and recoverability of vision-language-action (VLA) models in robotic manipulation tasks.
What happened
The article presents OGPO, a novel post-training framework for VLA models that addresses limitations in traditional trajectory-level imitation learning by optimizing for task-progress state reachability through semantic key state relabeling and goal-conditioned rewards.
Why it matters
For AI developers and robotics professionals, this approach enhances the reliability of VLA models in real-world scenarios where execution deviations from demonstrated trajectories are common, improving task completion rates in unstructured environments.
- OGPO addresses limitations of traditional imitation learning in VLA models
- Key-state reachability learning improves model robustness
- The approach requires no additional online data during post-training
- Benchmark results demonstrate consistent performance improvements
The bigger picture
The framework enables VLA models to achieve better performance and robustness without requiring additional online interaction, which could accelerate deployment in practical robotic systems requiring adaptability to unexpected situations.
OGPO provides a significant advancement in VLA post-training by shifting from trajectory imitation to goal-oriented state reachability, demonstrating improved performance and robustness in robotic manipulation tasks.
Technical details
- Offline goal-conditioned policy optimization
- Semantic key state relabeling
- Goal-conditioned process rewards
- Vision-language-action (VLA) models
- LIBERO and MetaWorld benchmarks
- LIBERO-ReAct mid-execution perturbation benchmark
FAQ
What problem does OGPO solve for VLA models?
OGPO addresses the limitation of traditional trajectory-level imitation learning, which biases VLA models toward demonstration-specific actions and limits their ability to recover from off-demonstration states.
How does OGPO differ from standard supervised fine-tuning?
Unlike standard supervised fine-tuning that focuses on replicating expert actions, OGPO optimizes for reaching semantic key states through goal-conditioned rewards, enabling better adaptability to execution deviations.
What benchmarks were used to evaluate OGPO?
The framework was tested on LIBERO, MetaWorld, LIBERO-Plus, and the proposed LIBERO-ReAct benchmark, which includes mid-execution perturbation scenarios.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - cs.AI on Oct 12, 2026.
Comments
Sign in to join the discussion.