港科大等团队提出WMPO：基于世界模型的VLA策略优化框架

Core Insights - The article introduces WMPO (World Model-based Policy Optimization), a framework developed by Hong Kong University of Science and Technology and ByteDance Seed team, which enhances sample efficiency, task performance, generalization ability, and lifelong learning through pixel-level video generation for VLA (Vision-Language-Action) models [5][25]. Research Background and Pain Points - Existing solutions struggle to balance scalability and effectiveness, with human intervention requiring continuous supervision and high costs for adapting simulators to diverse scenarios [4]. - Traditional latent space world models misalign with web-scale pre-trained visual features, failing to fully leverage pre-trained knowledge [4] [6]. Core Framework Design - WMPO's logic is based on generating trajectories in an "imagination" space using high-fidelity pixel-level world models, replacing real environment interactions and supporting stronger on-policy reinforcement learning [5][11]. - The iterative process follows "imagination trajectory generation → trajectory sampling evaluation → policy update" [5]. Key Modules - Generative World Model: Simulates dynamic changes between the robot and the environment, generating visual trajectories aligned with VLA pre-trained features [8]. - Lightweight Reward Model: Automatically assesses the success or failure of imagined trajectories, providing sparse reward signals to avoid complex reward shaping [9]. - On-Policy Policy Optimization (GRPO): Adapts Group Relative Policy Optimization for sparse reward scenarios, balancing stability and scalability [10]. Core Innovations - Pixel Space Priority: Directly generates trajectories in pixel space, perfectly matching VLA pre-trained visual features and maximizing the value of pre-trained knowledge [11]. - Trajectory Generation Logic: Predicts action blocks based on initial frames and language instructions, generating subsequent frames iteratively [12]. - Dynamic Sampling Strategy: Generates multiple imagined trajectories from the initial state, filtering out all-success or all-failure trajectories to ensure effective training samples [12]. Experimental Validation and Key Results - In simulation environments, WMPO outperformed baseline methods (GRPO, DPO) across four fine manipulation tasks, achieving an average success rate of 47.1% with a rollout budget of 128, and 57.6% with a budget of 1280, demonstrating superior sample efficiency [13][14]. - In real environments, WMPO achieved a success rate of 70% in a "block insertion" task, significantly higher than baseline strategies [15]. Emergent Behaviors - WMPO exhibits self-correcting capabilities, autonomously adjusting actions in response to failure states, unlike baseline strategies that continue erroneous actions until timeout [17]. Generalization Ability - WMPO demonstrated an average success rate of 29.6% in out-of-distribution scenarios, outperforming all baseline methods, indicating its learning of general operational skills rather than false visual cues [19][20]. Lifelong Learning - WMPO showed stable performance improvement through iterative collection of trajectories, while DPO struggled with instability and required more expert demonstrations [23]. Conclusion and Significance - WMPO establishes a new paradigm for VLA optimization by integrating world models with on-policy reinforcement learning, addressing high costs and low sample efficiency in real environment interactions. It enhances performance, generalization, and lifelong learning capabilities, paving the way for scalable applications in general robotic operations [25].