⌘K
Change language Switch ThemeSign In
Narrow Mode
你的模型真的会”举一反三”吗？RoboChallenge Table30 V2 正式发布，泛化时代开幕
量 量子位 @量子位的朋友们
One Sentence Summary
RoboChallenge Table30 V2 is officially released, providing a generalization evaluation standard for embodied AI models oriented towards large-scale, real-world native deployment through three major upgrades: tasks, evaluation, and systems.
Summary
The article details the release and core improvements of the embodied AI evaluation benchmark, RoboChallenge Table30 V2. Addressing the overfitting issues common in single-task fine-tuning of embodied AI models, Table30 V2 has been deeply reconstructed across three dimensions: tasks, evaluation, and systems. At the task level, it introduces high-difficulty scenarios such as soft object manipulation, tool usage, and bimanual collaboration. At the evaluation level, it shifts fully to a multi-task paradigm and introduces Zero-shot and Out-of-Distribution (OOD) generalization tests. At the system level, it achieves a 300% increase in throughput through hardware expansion. This benchmark will debut as the platform for the CVPR 2026 Workshop competition, aiming to drive embodied AI from single-point breakthroughs to full-dimensional generalization.
Main Points
* 1. Task complexity significantly increased, facing real-world challenges head-on.Adds 18 high-difficulty tasks, including soft object manipulation, tool usage, and bimanual collaboration, testing the model's spatial reasoning, physical common sense, and high-precision synchronous control capabilities. * 2. Evaluation protocol reconstructed, shifting from single-task fine-tuning to full-dimensional generalization.Introduces Zero-shot testing and Out-of-Distribution (OOD) generalization matrices to prevent overfitting optimization for specific tasks, advocating for the development of a single model with general understanding capabilities. * 3. Underlying infrastructure upgraded, significantly improving research iteration efficiency.System throughput increased by 3x, and 'completion time' is introduced as a key performance metric, compelling researchers to optimize strategy execution efficiency rather than blindly stacking inference time.
Metadata
AI Score
87
Website qbitai.com
Published At Today
Length 1851 words (about 8 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
!Image 6具身智能正站在一个关键的历史节点上。随着 VLA 与 WMA 模型爆发式增长，行业的核心痛点日益凸显：在单任务精调之外，具身智能模型是否真正具备通用泛化能力，还是仅仅陷入了单一任务的过拟合？
带着这一核心追问，RoboChallenge正式发布Table30 V2，它以“面向下一代模型的大规模真机原生泛化评测”为标准，从任务升级、评测升级到系统升级三个维度深度重构：通过引入更严苛的软体、工具使用及双臂协作任务，支持零样本与域外泛化测试，并实现3倍于往届的系统吞吐量。Table30 V2将为全球具身智能研究者打造了一个精准的“泛化标尺”与公平、开放的真机竞技场，同时其预览版将作为 RoboChallenge CVPR 2026 Workshop 竞赛的首秀上线发布。
一、任务升级：直面真实世界的复杂性与灵巧性
具身智能的终极目标是处理现实世界中不可预知的复杂场景。Table30 V2 的第一个核心改进，就是重构了任务集。我们在原有基础上，新增了 18 个全新的双臂灵巧操作任务，配合保留的 12 个经典任务，构建了一个包含 30 个高难度任务的综合评估体系。为什么要如此扩展任务难度？因为泛化能力的缺失往往源于对“简单任务”的过度拟合。
* 从硬到软的跨越：我们引入了对绳索、布料等软连续体物体的处理任务。这些物体形变无限、状态不固定，极大地挑战了模型的空间推理与自适应控制能力。 * 工具与空间的深度交互：新增任务要求机器人能够准确使用工具，并理解复杂的工具-物体空间关系。这不仅是对精度控制的考验，更是对模型物理常识的一次深度摸底。 * 双臂协作的必然性：在 Table30 V2 中，我们设计了大量需要双手高度协调的任务，要求模型在动态受限的环境下，实现高精度的同步控制。 * 硬件升级，验证跨平台能力：引入了新一代移动双臂操纵平台 DOS-W1（具有三角尖端夹具），并与经典的 Aloha系统并存，不仅降低了整体硬件的准入门槛，更通过双机型并行的评测机制，严谨地验证了模型在不同硬件配置下的鲁棒性。
二、评测升级：告别单点突破，迎接全维泛化
在具身智能领域，我们经常看到模型在单个任务上表现优异，却在稍换场景后即刻“翻车”。Table30 V2 通过重构学习与测试协议，彻底终结了“为了比赛而调参”的旧模式。
* 全面转向“多任务”范式：Table30 V2 将全面支持多任务模型，杜绝为每个任务单独训练一个专用模型的“作弊式”优化。我们要求参评者提交具备通用理解能力的单一模型，这与构建通用具身大模型的行业大趋势高度一致。 * 零样本（Zero-shot）测试的引入：这是本次升级的重头戏。我们引入了物体级和环境级两个层面的零样本测试。模型必须在未见过的物体、未见过的背景、甚至动态变化的干扰（如桌面高度微调）下，展示出其真正的理解力，而非单纯的记忆能力。
分层次的泛化矩阵（In-Domain vs. Out-of-Domain）：从传统的域内评估，升级为包含域外（OOD）的高阶测试。我们不仅测试模型在预期之内的表现，更通过动态调整测试环境（如将桌面换成沙发等不可控变化），挑战模型的边界。这不再是一场简单的分数比拼，而是一次对模型“智能本质”的压力测试。
三、系统升级：吞吐量提升300%，打造快节奏反馈循环
对于算法迭代而言，速度就是生命线。为了让科研人员不再因为“等待评估结果”而浪费宝贵的算力，我们对整个底层基础设施进行了彻底的扩容与优化。
* 300% 的吞吐量飞跃：通过大量增购主流机器人硬件，并优化调度算法，我们实现了系统吞吐能力的巨大提升。 * 任务准备的“极速化”：我们放宽了像素级的严苛初始状态要求，转向更符合现实的“粗略对齐”，大幅降低了任务准备的空转时间，确保科研团队能够实时获取反馈，快速迭代模型。 * “时间消耗”作为性能指标：我们在排行榜中新增了“完成时间（Time to Complete）”作为关键评分维度，倒逼研究者优化策略的执行效率，而非盲目堆叠推理时间来换取成功率。
诚邀报名：RoboChallenge CVPR 2026 Workshop竞赛正式开启
Table30 V2 的发布，不仅是一个基准测试的诞生，更是一次对未来具身智能方向的指引。我们诚挚邀请全球科研团队、实验室及个人开发者，参与 RoboChallenge CVPR 2026 Workshop 竞赛。本次竞赛将作为 Table30 V2 预览版首秀，参赛者有机会在我们的真实机器人集群上验证自己的模型，与全球顶尖的具身智能算法同台竞技。
竞赛结束后，我们将持续面向全球研究者开放评测基础设施，让每一个有想法的团队都能在真实机器上验证自己的模型。泛化，是具身智能的下一座山峰。Table30 V2，邀你同攀。
详情敬请关注 RoboChallenge 官方渠道：https://robochallenge.cn/competition。
*本文系量子位获授权刊载，观点仅为原作者所有。
量 量子位 @量子位的朋友们
One Sentence Summary
RoboChallenge Table30 V2 is officially released, providing a generalization evaluation standard for embodied AI models oriented towards large-scale, real-world native deployment through three major upgrades: tasks, evaluation, and systems.
Summary
The article details the release and core improvements of the embodied AI evaluation benchmark, RoboChallenge Table30 V2. Addressing the overfitting issues common in single-task fine-tuning of embodied AI models, Table30 V2 has been deeply reconstructed across three dimensions: tasks, evaluation, and systems. At the task level, it introduces high-difficulty scenarios such as soft object manipulation, tool usage, and bimanual collaboration. At the evaluation level, it shifts fully to a multi-task paradigm and introduces Zero-shot and Out-of-Distribution (OOD) generalization tests. At the system level, it achieves a 300% increase in throughput through hardware expansion. This benchmark will debut as the platform for the CVPR 2026 Workshop competition, aiming to drive embodied AI from single-point breakthroughs to full-dimensional generalization.
Main Points
* 1. Task complexity significantly increased, facing real-world challenges head-on.
Adds 18 high-difficulty tasks, including soft object manipulation, tool usage, and bimanual collaboration, testing the model's spatial reasoning, physical common sense, and high-precision synchronous control capabilities.
* 2. Evaluation protocol reconstructed, shifting from single-task fine-tuning to full-dimensional generalization.
Introduces Zero-shot testing and Out-of-Distribution (OOD) generalization matrices to prevent overfitting optimization for specific tasks, advocating for the development of a single model with general understanding capabilities.
* 3. Underlying infrastructure upgraded, significantly improving research iteration efficiency.
System throughput increased by 3x, and 'completion time' is introduced as a key performance metric, compelling researchers to optimize strategy execution efficiency rather than blindly stacking inference time.
Key Quotes
* The lack of generalization capability often stems from overfitting to 'simple tasks'. * By reconstructing learning and testing protocols, Table30 V2 has completely ended the old model of 'tuning parameters just for the competition'. * Generalization is the next peak for embodied AI. Table30 V2 invites you to climb it together.
AI Score
87
Website qbitai.com
Published At Today
Length 1851 words (about 8 min)
Tags
Embodied AI
RoboChallenge
Table30 V2
Generalization Evaluation
Zero-shot Learning
Related Articles
* 150: Year-End AI Review: From Models to Applications, Technology to Business Wars, Grasping the Thread of Meaning in the Torrent * After Large Models: Rewriting the Division of Labor Between Humans and Machines | A 10,000-Word Roundtable Transcript * Zhipu GLM-5 Technical Deep Dive: Full Compatibility with Domestic Chips Like Huawei Sparks Global Discussion and asynchronous reinforcement learning infrastructure, while completing full-stack adaptation for domestic chips.") * The First Three Years of All-in AI | A Conversation with Zhang Jinjian, Partner at Vitalbridge * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * Vol.94 | A Conversation with Zibianliang Robotics: Three Bold Assertions on Robot Data * AI Starts to "Take Action", Alibaba's Qwen Leads the World * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * After Topping Open-Source Rankings with its Programming LLM, the Zhipu GLM Team Faced a 3-Hour Questioning Session * 61. Minimax Goes Public: Three Years of Evolution in the LLM Landscape | A Conversation with Chen Yu from Yunqi Partners HomeArticlesPodcastsVideosTweets