港中文最新！无需微调即可部署VLA模型

Core Insights - The article introduces VLA-Pilot, a plug-and-play inference-time strategy that enhances the deployment of pre-trained VLA models in real-world robotic tasks without requiring additional fine-tuning or data collection [4][6][35] - VLA-Pilot significantly improves the success rate of pre-trained VLA strategies across diverse tasks and robot forms, demonstrating robust zero-shot generalization capabilities [4][6] Current Issues - Pre-trained VLA strategies often experience performance degradation when deployed in downstream tasks, which can be mitigated through fine-tuning, but this approach is costly and impractical in real-world scenarios [4][5] - Existing methods for guiding pre-trained strategies during inference have limitations, including the need for additional training and reliance on fixed candidate action sets, which may not align with task contexts [5][6] Innovations and Methodology - VLA-Pilot utilizes a multi-modal large language model (MLLM) as an open-world validator to enhance generalization and employs an evolutionary diffusion process for action optimization, improving task alignment [6][10] - The method includes an embodied policy steering chain (EPS-CoT) module that infers guiding target rewards from task contexts without requiring task-specific training [11][12] - An iterative guiding optimization mechanism is integrated to ensure closed-loop corrections, enhancing the precision and contextual relevance of the guiding process [20][21] Experimental Analysis - VLA-Pilot was evaluated using a dual-arm system, demonstrating superior performance compared to six baseline methods in both in-distribution and out-of-distribution tasks [23][24] - The experiments included six downstream tasks, with metrics such as operation success rate (MSR) and guiding objective alignment (SOA) used to assess performance [26][27] - Results showed that VLA-Pilot outperformed all baseline methods in in-distribution tasks and exhibited robust generalization capabilities in out-of-distribution tasks [28][31] Comparative Performance - In in-distribution tasks, VLA-Pilot achieved an overall MSR of 0.62 and SOA of 0.73, outperforming all baseline methods [30] - For out-of-distribution tasks, VLA-Pilot demonstrated a significant success rate of 0.50, indicating strong adaptability to unseen scenarios [32] Conclusion - VLA-Pilot effectively maximizes the utility of existing VLA models during inference, providing a scalable and data-efficient solution for robotic operations [35]