以下文章来源于具身智能之心 ，作者具身智能之心 >> 点击进入→ 具身智能之心 技术交流群 更多VLA与RL实战项目，欢迎加入国内首个工业级VLA实战课程 ： 具身VLA实战与求职教程来啦～ 。 原文链接：https://vedder.io/misc/state_of_robot_learning_dec_2025.html 这次来学习一下 PI 内部人员写的 blog，介绍了很多 robot learning 的现状，而且都是一线的真正经验，很多在一线的同学应该深有感 触，说了很多实话，质量很高，值得精读和学习。不管是对 IL DAgger RL 的看法都是很一手的经验。 接下来请享受这份知识 具身智能之心 . 与世界交互，更进一步。具身智能之心是国内具身与机器人领域的专业技术平台，集企业咨询、在线教育、展会服务、线下培 训、硬件研发、技术方案为一体。 点击下方 卡片 ，关注" 具身智能 之心 "公众号 编辑丨 具身智能之心 本文只做学术分享，如有侵权，联系删文 基本上，目前（2025 年 12 月）所有机器人学习系统都是纯粹的行为克隆（BC，也称模仿学习）系统。人类提供（接近）最优的任务演 示，机器学习模 ...

点击下方 卡片 ，关注" 具身智能 之心 "公众号 编辑丨 具身智能之心 本文只做学术分享，如有侵权，联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多VLA与RL实战项目，欢迎加入国内首个工业级VLA实战课程 ： 具身VLA实战与求职教程来啦～ 。 原文链接：https://vedder.io/misc/state_of_robot_learning_dec_2025.html 这次来学习一下 PI 内部人员写的 blog，介绍了很多 robot learning 的现状，而且都是一线的真正经验，很多在一线的同学应该深有感触，说了很多实话，质量很 高，值的精读和学习。不管是对 IL DAgger RL 的看法都是很一手的经验。 接下来请享受这份知识 基本上，目前（2025 年 12 月）所有机器人学习系统都是纯粹的行为克隆（BC，也称模仿学习）系统。人类提供（接近）最优的任务演示，机器学习模型则尝试模 仿这些动作。形式上，策略训练采用监督式方法——给定机器人的状态 （例如摄像头图像、机器人关节角度以及可能的任务描述文本），policy 预测已演示的动作 a 通常是一个动作片段（action chun ...

Core Viewpoint - The article discusses the development and capabilities of Vision-Language-Action (VLA) models, emphasizing their application in general robot control and the integration of multimodal understanding for enhanced performance in complex tasks [6][8]. Group 1: VLA Model Development - The VLA models are designed to evolve continuously, integrating semantic reasoning from open-world scenarios to execute dexterous actions in real environments [6]. - A unified framework combining perception and action is established, addressing challenges such as flexible objects and fine manipulation tasks to enhance general dexterity [6]. Group 2: Model Capabilities and Enhancements - The article highlights the expansion of VLA model capabilities, focusing on improving generalization abilities through the incorporation of large model priors and semantic reasoning [8]. - The introduction of Spec-VLA, a framework specifically designed for accelerating inference in VLA models, is noted as a significant advancement [8]. Group 3: Expert Insights and Additional Content - The live session features insights from industry experts, including the head of the embodied foundation model at Midea, discussing the design challenges of dexterous hands and their role in closing the "hand-eye-brain" perception loop [7]. - Additional content is available on the knowledge platform "Embodied Intelligence Heart," providing in-depth analysis and technical details related to VLA models and their applications [10].

Core Viewpoint - The article presents AutoOcc, an innovative framework for automatic open-ended 3D semantic occupancy annotation that surpasses existing methods without requiring human labeling, showcasing excellent generalization capabilities [5][11][26]. Summary by Sections Introduction - AutoOcc is developed by the VDIG laboratory at Peking University, led by researchers Zhou Xiaoyu and Wang Yongtao, and has been recognized in top conferences and competitions in the computer vision field [2][4]. Problem Statement - The challenge of generating accurate and complete semantic occupancy annotations from raw sensor data at low cost remains significant in the fields of autonomous driving and embodied intelligence [5][8]. Methodology - AutoOcc utilizes a vision-language model (VLM) to create semantic attention maps for scene description and dynamically expands the semantic list, while a self-estimating optical flow module identifies and processes dynamic objects in temporal rendering [5][11][17]. Key Innovations - The framework introduces a 3D Gaussian representation (VL-GS) that effectively models complete 3D geometry and semantics in driving scenarios, demonstrating superior representation efficiency, accuracy, and perception capabilities [6][17]. Experimental Results - Extensive experiments indicate that AutoOcc outperforms existing automated 3D semantic occupancy annotation methods and exhibits remarkable zero-shot generalization across datasets [7][21][22]. Comparison with Existing Methods - AutoOcc is compared with traditional methods that rely on human labeling and extensive post-processing, highlighting its speed and open-ended semantic annotation capabilities [14][21]. Performance Metrics - The framework shows significant advantages in terms of robustness and open semantic labeling ability, achieving state-of-the-art performance in both specific semantic categories and across datasets [20][21]. Efficiency Evaluation - AutoOcc demonstrates a notable reduction in computational costs while enhancing annotation performance, achieving a balance between efficiency and flexibility without relying on human annotations [24][25]. Conclusion - The article concludes that AutoOcc represents a significant advancement in automated open semantic 3D occupancy annotation, integrating visual language model guidance with differentiable 3D Gaussian techniques [26].

Research Background and Core Issues - The development of embodied intelligence has led to the integration of robots into daily life as human assistants, necessitating their ability to interpret high-level instructions, perceive dynamic environments, and adjust plans in real-time [3] - Vision-Language Models (VLMs) have emerged as a significant direction for robot task planning, but existing methods exhibit limitations in three areas: insufficient interactive exploration capabilities, limited perception accuracy, and poor planning adaptability [6] Proposed Framework - The ExploreVLM framework is introduced, which integrates perception, planning, and execution verification through a closed-loop design to address the identified limitations [5] Core Framework Design - ExploreVLM operates on a "perception-planning-execution-verification" closed-loop model, which includes: 1. Insufficient interactive exploration capabilities for scenarios requiring active information retrieval [6] 2. Limited perception accuracy in capturing object spatial relationships and dynamic changes [6] 3. Poor planning adaptability, primarily relying on open-loop static planning, which can fail in complex environments [6] Key Module Analysis 1. **Goal-Centric Spatial Relation Graph (Scene Perception)** - Constructs a structured graph representation to support complex reasoning, extracting object categories, attributes, and spatial relationships from initial RGB images and task goals [8] - A two-stage planner generates sub-goals and action sequences for exploration and completion phases, optimizing through self-reflection [8] - The execution validator compares pre- and post-execution states to generate feedback and dynamically adjust plans until task completion [8] 2. **Dual-Stage Self-Reflective Planner** - Designed to separate the needs for "unknown information exploration" and "goal achievement," employing a self-reflection mechanism to correct plans and address logical errors [10] - The exploration phase generates sub-goals for information retrieval, while the completion phase generates action sequences based on exploration results [10] 3. **Execution Validator** - Implements a step-by-step validation mechanism to ensure real-time feedback integration into the closed loop, supporting dynamic adjustments [14] Experimental Validation 1. **Experimental Setup** - Conducted on a real robot platform with five tasks of increasing complexity, comparing against baseline methods ReplanVLM and VILA, with a 50% action failure rate introduced to test robustness [15] 2. **Core Results** - ExploreVLM achieved an average success rate of 94%, significantly outperforming ReplanVLM (22%) and VILA (30%) [16] - The framework demonstrated effective action validation and logical consistency checks, ensuring task goals were met [17] 3. **Ablation Studies** - Performance dropped significantly when core modules were removed, highlighting the importance of the collaborative function of the three modules [19] Comparison with Related Work - ExploreVLM addresses the limitations of existing methods through structured perception, dual-stage planning, and stepwise closure, enhancing task execution and adaptability [20]

Core Insights - The article discusses the development of NavigScene, a novel dataset aimed at bridging the gap between local perception and global navigation in autonomous driving systems, enhancing their reasoning and planning capabilities [2][12][14]. Group 1: Overview of NavigScene - NavigScene is designed to integrate local sensor data with global navigation context, addressing the limitations of existing autonomous driving models that primarily rely on immediate visual information [5][9]. - The dataset includes two subsets: NavigScene-nuScenes and NavigScene-NAVSIM, which provide paired data to facilitate comprehensive scene understanding and decision-making [9][14]. Group 2: Methodologies - Three complementary paradigms are proposed to leverage NavigScene: 1. Navigation-guided reasoning (NSFT) enhances visual-language models by incorporating navigation context [10][19]. 2. Navigation-guided preference optimization (NPO) improves generalization in new scenarios through reinforcement learning [24][26]. 3. Navigation-guided visual-language-action (NVLA) model integrates navigation guidance with traditional driving models for better performance [27][28]. Group 3: Experimental Results - Experiments demonstrate that integrating global navigation knowledge significantly improves the performance of autonomous driving systems in tasks such as perception, prediction, and planning [12][34][39]. - The results indicate that models trained with NavigScene outperform baseline models across various metrics, including BLEU-4, METEOR, and CIDEr, showcasing enhanced reasoning capabilities [32][34]. Group 4: Practical Implications - The integration of NavigScene allows autonomous systems to make more informed decisions in complex driving environments, leading to improved safety and reliability [12][42]. - The findings highlight the importance of incorporating beyond-visual-range (BVR) knowledge for effective navigation and planning in autonomous driving applications [8][12].

Core Insights - The article discusses the introduction of V-Triune, a unified reinforcement learning system by MiniMax that enhances visual-language models (VLM) for both visual reasoning and perception tasks in a single training process [2][4][5]. Group 1: V-Triune Overview - V-Triune consists of three complementary components: Sample-Level Data Formatting, Verifier-Level Reward Computation, and Source-Level Metric Monitoring, which work together to handle diverse tasks [3][8]. - The system utilizes a novel dynamic IoU reward mechanism that provides adaptive feedback for perception tasks, leading to performance improvements in reasoning and perception tasks [3][4]. Group 2: Performance Improvements - Orsta, the model generated by V-Triune, achieved significant performance gains in the MEGA-Bench Core benchmark, with improvements ranging from +2.1 to +14.1 across different model variants [4][49]. - The model's training on diverse datasets covering various visual reasoning and perception tasks has contributed to its broad capabilities [3][49]. Group 3: Sample-Level Data Formatting - MiniMax addresses the challenge of different tasks requiring distinct reward types and configurations by defining rewards at the sample level, allowing for dynamic routing and fine-grained weighting during training [9][13][16]. - This design enables seamless integration of diverse datasets into a unified training process while allowing for flexible and scalable reward control [16]. Group 4: Verifier-Level Reward Computation - MiniMax employs an independent, asynchronous reward server for generating reinforcement learning signals, enhancing modularity and scalability [17][19]. - The architecture allows for easy addition of new tasks or updates to reward logic without modifying the core training process [20]. Group 5: Source-Level Metric Monitoring - The Source-Level Metric Monitoring strategy records key performance indicators by data source for each training batch, facilitating targeted debugging and insights into the interactions between different data sources [21][24]. - Key monitored metrics include dynamic IoU rewards, perception task IoU/mAP, response length, and reflection rate, all tracked continuously by data source [24][22]. Group 6: Dynamic IoU Reward Strategy - The dynamic IoU reward strategy adjusts the IoU threshold during training to balance learning efficiency and final accuracy, starting with a relaxed threshold and progressively tightening it [26][25]. - This approach aims to guide the model's learning process smoothly while ensuring high performance in the later stages of training [26]. Group 7: Training Methodology - MiniMax's V-Triune supports scalable data, tasks, validators, and metrics systems, but early experiments indicated that joint training could lead to instability [28][29]. - To address this, MiniMax implemented targeted adjustments, including freezing ViT parameters to prevent gradient explosion and managing memory during large-scale training [34][35]. Group 8: Experimental Results - MiniMax conducted experiments using Qwen2.5-VL-7B-Instruct and Qwen2.5-VL-32B-Instruct as base models, achieving a dataset comprising 20,600 perception samples and 27,100 reasoning samples [46]. - The results indicate that V-Triune significantly enhances performance in reasoning and perception tasks, particularly in areas with rich training data [49][55]. Group 9: Conclusion - Overall, MiniMax's findings suggest that reinforcement learning can effectively enhance visual reasoning and perception capabilities within a unified framework, demonstrating continuous performance improvements across various tasks [55][56].