Core Viewpoint - The article discusses the evolution of autonomous driving technology, emphasizing the need for a unified perspective on various paradigms, including end-to-end (E2E), VLM-centric, and hybrid approaches, to enhance understanding and performance in complex driving scenarios [2][4][14]. Group 1: Introduction and Background - Traditional modular approaches in autonomous driving have led to information loss and error accumulation due to task fragmentation, prompting a shift towards data-driven end-to-end architectures [5][10]. - The article introduces a comprehensive review titled "Survey of General End-to-End Autonomous Driving: A Unified Perspective," which aims to bridge the gap in understanding between different paradigms [3][4]. Group 2: Paradigms of Autonomous Driving - General End-to-End (GE2E) is defined as any model that processes raw sensor inputs into planning trajectories or control actions, regardless of whether it includes visual-language models (VLM) [4][14]. - The three main paradigms unified under GE2E are: - Traditional End-to-End (Conventional E2E), which relies on structured scene representation for precise trajectory planning [9][17]. - VLM-centric End-to-End, which utilizes pre-trained visual-language models to enhance generalization and reasoning capabilities in complex scenarios [11][33]. - Hybrid End-to-End, which combines the strengths of both traditional and VLM-centric approaches to balance high-level semantic understanding with low-level control precision [12][39]. Group 3: Performance Comparison - In open-loop performance tests, the hybrid paradigm outperformed others, demonstrating the importance of world knowledge in handling long-tail scenarios [54]. - Traditional E2E methods still dominate in numerical trajectory prediction accuracy, indicating their robustness in structured environments [54]. - In closed-loop performance, traditional methods maintain a stronghold, particularly in complex driving tasks, while VLA methods show potential but require further refinement in fine-grained trajectory control [55][56]. Group 4: Data and Learning Strategies - The evolution of datasets from geometric annotations to semantic-rich datasets is crucial for training models capable of logical reasoning and understanding complex traffic contexts [46][48]. - The introduction of Chain of Thought (CoT) annotations in datasets supports advanced reasoning tasks, moving beyond simple input-output mappings [47]. Group 5: Model Architecture and Details - The article provides a detailed comparison of mainstream model architectures, including their inputs, backbone networks, intermediate tasks, and output forms, to clarify the distinctions among different paradigms [57].

Core Insights - The article presents DeepThinkVLA, a new model that addresses the challenges in the visual-language-action (VLA) domain by integrating a mixed attention decoder and a two-stage training pipeline, achieving a task success rate of 97.0% on the LIBERO benchmark, setting a new performance standard for VLA models [2][14]. Group 1: Model Architecture - DeepThinkVLA resolves the "modal conflict" between reasoning and action by employing a mixed attention mechanism that allows for efficient processing of both modalities within a single decoder [4][10]. - The model features a dynamic switching mechanism between causal attention for reasoning generation and bidirectional attention for action generation, significantly reducing inference latency and enhancing performance [4][10]. Group 2: Training Methodology - The training process consists of a two-stage pipeline combining supervised fine-tuning (SFT) and reinforcement learning (RL), which enhances the model's reasoning capabilities while ensuring effective action execution [6][8]. - The SFT phase focuses on building foundational reasoning skills through a carefully designed data augmentation pipeline, resulting in a dataset of 273,465 annotated frames [10][12]. Group 3: Innovations and Mechanisms - Two key innovations are highlighted: the probabilistic decomposition of reasoning and action, and an error recovery mechanism that allows the model to self-correct during execution [10][11]. - The reward design incorporates task-success rewards and format regularization rewards, focusing on the final success of tasks while minimizing interference from intermediate reasoning semantics [11][12]. Group 4: Performance Evaluation - DeepThinkVLA outperforms existing models across various tasks, achieving an average success rate of 97.0%, with specific task success rates of 99.0% for Object tasks and 96.4% for Goal tasks [14][15]. - The model demonstrates superior robustness compared to top autoregressive models, showcasing its effectiveness in complex robotic operations [15][16]. Group 5: Future Directions - Future enhancements may include integrating additional sensory data, expanding to more complex collaborative tasks, optimizing efficiency, and constructing larger datasets to improve model generalization [23][24].