Beyond Success: Refining Elegant Robot Manipulation from Mixed-Quality Data via Just-in-Time Intervention

2025-11-27

AI Summary

To address unstable policy execution in Vision-Language-Action (VLA) models for robotic manipulation—caused by heterogeneous quality of human demonstration data—this paper proposes a decoupled, training-free real-time intervention framework. Our approach comprises three key contributions: (1) introducing LIBERO-Elegant, the first benchmark explicitly evaluating manipulation elegance; (2) formulating implicit task constraint modeling to decouple aesthetic action quality from task success; and (3) training an elegance evaluator via offline-calibrated Q-learning, enabling dynamic assessment and refinement of actions at critical decision points. Experiments demonstrate substantial improvements in execution quality across both simulation and real-world settings, with strong generalization to unseen tasks. The framework shifts control granularity from binary “task completion” to fine-grained “elegant completion,” advancing the fidelity and robustness of VLA-based robotic manipulation.

Technology Category

• Intelligent Robots: Manipulation
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Computer Vision: Visual Reasoning & Symbolic Representations

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Vision-Language-Action (VLA) models have enabled notable progress in general-purpose robotic manipulation, yet their learned policies often exhibit variable execution quality. We attribute this variability to the mixed-quality nature of human demonstrations, where the implicit principles that govern how actions should be carried out are only partially satisfied. To address this challenge, we introduce the LIBERO-Elegant benchmark with explicit criteria for evaluating execution quality. Using these criteria, we develop a decoupled refinement framework that improves execution quality without modifying or retraining the base VLA policy. We formalize Elegant Execution as the satisfaction of Implicit Task Constraints (ITCs) and train an Elegance Critic via offline Calibrated Q-Learning to estimate the expected quality of candidate actions. At inference time, a Just-in-Time Intervention (JITI) mechanism monitors critic confidence and intervenes only at decision-critical moments, providing selective, on-demand refinement. Experiments on LIBERO-Elegant and real-world manipulation tasks show that the learned Elegance Critic substantially improves execution quality, even on unseen tasks. The proposed model enables robotic control that values not only whether tasks succeed, but also how they are performed.

Problem

• Improving robot execution quality from mixed-quality human demonstrations
• Developing a decoupled framework to refine VLA policies without retraining
• Enhancing robotic manipulation by satisfying implicit task constraints elegantly

Innovation

• Decoupled refinement framework for VLA policies
• Elegance Critic trained via offline Calibrated Q-Learning
• Just-in-Time Intervention mechanism for selective refinement