⌘K
Change language Switch ThemeSign In
Narrow Mode
具身 Scaling Law 押对了！独角兽新品 1 小时学会新任务，重复 1800 次成功率 99%
量 量子位 @克雷西
One Sentence Summary
Embodied AI unicorn Generalist releases the Gen-1 model, validating the Embodied Scaling Law through pre-training on large-scale human activity data, significantly improving robot success rates and efficiency in complex tasks.
Summary
This article introduces the Gen-1 model, the latest research breakthrough from embodied AI startup Generalist. In fine-motor tasks such as packaging phones and folding cardboard boxes, the model increased robot success rates from 64% to 99% and improved efficiency by 3x. The core breakthrough of Gen-1 lies in its 'de-robotized' pre-training scheme, which utilizes millions of human activity records captured by low-cost wearable devices rather than expensive teleoperation data. This allows the model to grasp physical causality before ever touching a robotic arm. Technically, the team introduced a Paged Attention mechanism specifically designed for physical interaction to ensure millisecond-level response times, and adopted a Harmonic Reasoning system for multi-scale dynamic adjustment. The success of Gen-1 proves the validity of the Scaling Law in the field of embodied AI, enabling the robot to demonstrate improvisational intelligence when handling unexpected situations.
Main Points
* 1. The Gen-1 model significantly improves the success rate and efficiency of robots performing fine-motor tasks.In tasks like folding cardboard boxes, the success rate jumped from 64% to 99%, and completion time was reduced from 34 seconds to 12.1 seconds, demonstrating immense industrial application potential. * 2. Adopts a 'de-robotized' large-scale pre-training scheme to bypass data bottlenecks.By capturing millions of human activity records via low-cost wearable devices, the model gains insights into space, time, and physical causality from a human perspective before ever interacting with a robotic arm, increasing learning efficiency by 10x. * 3. Introduces Paged Attention and Harmonic Reasoning to optimize inference performance.Paged Attention solves computational scheduling challenges under PB-level data streams, ensuring instantaneous action command execution; Harmonic Reasoning allows the model to perform multi-scale weight adjustments during complex dynamic tasks. * 4. Validates the applicability of the Scaling Law in the field of embodied AI.The investment in large-scale data and compute has led the robot to an 'aha moment' in understanding physical laws, granting it the improvisational problem-solving ability to handle unexpected situations (such as misaligned parts).
Metadata
AI Score
88
Website qbitai.com
Published At Today
Length 1865 words (about 8 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
##### 克雷西 发自 凹非寺
量子位 | 公众号 QbitAI
机器人也开始内卷了，一位表现极其离谱的“新员工”，直接拉高了机器人的“就业门槛”。
具身智能独角兽Generalist，刚刚推出了最新的研究成果——新模型Gen-1。
在包装手机和折叠纸箱这些精细活儿上，它把机器人的成功率从64%硬生生拉到了99%，几乎告别了手残职业病。
!Image 8: img 以前折叠一个标准纸箱需要慢悠悠地磨掉34秒，现在GEN-1仅用12.1秒就能完成，效率直接开启了3倍速模式。
!Image 9: 图片 而且，GEN-1的表现，也用实际表现验证了机器人领域的Scaling Law。
机器人模型卷出新高度
GEN-1上岗后的表现简直像是在倍速播放，而且即便面对维护扫地机器人200次这种枯燥任务，它也能稳如泰山。
!Image 10: img 甚至连续装箱1800次，也能从从容容游刃有余。
!Image 11: img 更离谱的是它处理突发状况的脑回路。
如果零件在流水线上被意外撞歪了，它绝不会傻站在那儿报错，会自己切换抓取角度，甚至动用两只手配合着把活干完。
!Image 12: img 这种靠直觉解决问题的即兴智能，让它在处理乱七八糟的杂物时表现得像个干了十年的老师傅，那种死读程序的铁疙瘩僵硬感彻底消失了。
用人类活动记录训练机器人
为了让GEN-1具备使机器人变身“全能打工人”的能力，研发团队对数据处理架构进行了重写。
他们没有死磕昂贵且难以扩展的机器人遥操作数据这条老路，转而通过低成本穿戴设备捕捉了数百万项人类活动记录，让AI像看电影一样预习物理世界的潜规则。
这种“去机器人化”的预训练方案巧妙绕过了数据规模的瓶颈，让基础模型在接触机械臂之前，就已经从人类视角洞察了空间、时间与物理因果。
这种基于50万小时高保真物理交互数据集练就的底座，让它的学习效率直接起飞，达到了前代模型的10倍。
哪怕是面对从未见过的奇怪任务或陌生的机器身体，给GEN-1一个小时的实机演示，也能让它火速入职。
另外，为了让机器人的动作不再卡成PPT，以及实现实时操控，研发团队还在推理端祭出了两项关键技术。
首先是专门为物理世界打造的分页注意力（Paged Attention）机制。
在处理PB级别的物理交互数据流时，传统的内存管理方式容易导致计算资源分配不均，进而产生响应延迟。
Paged Attention通过更高效地调度计算资源，解决了动作指令发射时的调度难题，确保每一个动作指令都能在毫秒级的时间维度内即时发射，让AI的反应速度能跟上现实世界的物理节奏。
配合演进的还有一套Harmonic Reasoning系统。它作为推理层面的核心组件，改变了以往单一路径预测动作的死板模式。
它允许模型在输出指令时通过多尺度的动态调节来引导权重，使其在执行折叠纸箱或包装手机等复杂动态任务时，能够展现出超越单一模型权重分布的性能上限。
研发团队为此投入数月时间优化训练稳定性，并编写了大量自定义内核来压榨硬件算力的极限。
GEN-1的性能跨越，证明了Scaling Law在物理世界依然有效——只要喂够了数据和算力，机器人的脑子也会产生“开窍”时刻。
通过大规模预训练，机器人不再生硬地模仿动作序列，自己悟出了空间、时间和因果关系的规律，感知到了物体之间的相互影响。
有了直觉之后，机器人干活就开始带点“灵性”。当任务中途出现没见过的阻碍，它会自发尝试一些教学大纲以外的操作，比如发现东西塞不进去时会像人一样晃晃袋子。
这种即兴解题的能力源于它真正理解了“动作会导致后果”的逻辑。
即使现场零件被意外撞歪，它也能凭直觉找回节奏，不需要人类像保姆一样每一步都盯着纠错。
这种在真实世界摔打出来的经验，让原本悬浮在百科全书里的抽象文字变成了实打实的行动力。
研发团队通过对齐技术，给这种即兴天赋装上了“导航仪”，确保机器人“临场发挥”的动作依然会严丝合缝地待在用户设定的规范里。
这种进化，让机器人从一个只能按部就班的机器，变成了一个真正懂物理常识、能独立处理复杂局面的“职场老手”。
DeepMind大牛创业成果
GEN-1的底层逻辑，源于资深团队在机器人领域的长期积累，创始人Pete Florence的技术背景，为这一方案提供了深厚的理论底色。
!Image 13: img 他曾任Google DeepMind高级研究科学家，通过Dense Object Nets等工作探索了视觉引导下机器人从感知到动作的端到端学习路径。
在谷歌PaLM团队工作期间，他作为核心力量参与并主导了PaLM-E、RT-2等多个具备代际跨越意义的机器人项目。
2024年，Pete Florence离开谷歌并创立了Generalist。
即便在他离职后的2025年3月，DeepMind在发布的Gemini Robotics论文中依然四次引用了他参与合著的研究。
参考链接：
https://generalistai.com/blog/apr-02-2026-GEN-1
量 量子位 @克雷西
One Sentence Summary
Embodied AI unicorn Generalist releases the Gen-1 model, validating the Embodied Scaling Law through pre-training on large-scale human activity data, significantly improving robot success rates and efficiency in complex tasks.
Summary
This article introduces the Gen-1 model, the latest research breakthrough from embodied AI startup Generalist. In fine-motor tasks such as packaging phones and folding cardboard boxes, the model increased robot success rates from 64% to 99% and improved efficiency by 3x. The core breakthrough of Gen-1 lies in its 'de-robotized' pre-training scheme, which utilizes millions of human activity records captured by low-cost wearable devices rather than expensive teleoperation data. This allows the model to grasp physical causality before ever touching a robotic arm. Technically, the team introduced a Paged Attention mechanism specifically designed for physical interaction to ensure millisecond-level response times, and adopted a Harmonic Reasoning system for multi-scale dynamic adjustment. The success of Gen-1 proves the validity of the Scaling Law in the field of embodied AI, enabling the robot to demonstrate improvisational intelligence when handling unexpected situations.
Main Points
* 1. The Gen-1 model significantly improves the success rate and efficiency of robots performing fine-motor tasks.
In tasks like folding cardboard boxes, the success rate jumped from 64% to 99%, and completion time was reduced from 34 seconds to 12.1 seconds, demonstrating immense industrial application potential.
* 2. Adopts a 'de-robotized' large-scale pre-training scheme to bypass data bottlenecks.
By capturing millions of human activity records via low-cost wearable devices, the model gains insights into space, time, and physical causality from a human perspective before ever interacting with a robotic arm, increasing learning efficiency by 10x.
* 3. Introduces Paged Attention and Harmonic Reasoning to optimize inference performance.
Paged Attention solves computational scheduling challenges under PB-level data streams, ensuring instantaneous action command execution; Harmonic Reasoning allows the model to perform multi-scale weight adjustments during complex dynamic tasks.
* 4. Validates the applicability of the Scaling Law in the field of embodied AI.
The investment in large-scale data and compute has led the robot to an 'aha moment' in understanding physical laws, granting it the improvisational problem-solving ability to handle unexpected situations (such as misaligned parts).
Key Quotes
* This 'de-robotized' pre-training scheme cleverly bypasses the data scale bottleneck, allowing the foundation model to gain insights into space, time, and physical causality from a human perspective before ever touching a robotic arm. * The performance leap of GEN-1 proves that the Scaling Law remains valid in the physical world—as long as you feed it enough data and compute, the robot's brain will also have its 'aha moments'. * This ability for improvisational problem-solving stems from its genuine understanding of the logic that 'actions lead to consequences'.
AI Score
88
Website qbitai.com
Published At Today
Length 1865 words (about 8 min)
Tags
Embodied AI
Scaling Law
Gen-1
Generalist
Robotics
Related Articles
* 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.") * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * After Reading 200 Papers: Analyzing DeepMind, Meta, DeepSeek - What AGI Narratives Are Chinese and American Giants Describing? | 2025 AI Annual Review * 150: Year-End AI Review: From Models to Applications, Technology to Business Wars, Grasping the Thread of Meaning in the Torrent * The First Three Years of All-in AI | A Conversation with Zhang Jinjian, Partner at Vitalbridge * AI Starts to "Take Action", Alibaba's Qwen Leads the World * New Year 40,000-Word Dialogue with Elon Musk: Chinese Chips, Singularity, Age of Abundance and Humanity's Ultimate Evolution | Infographic + Video | Full Text Included * 61. Minimax Goes Public: Three Years of Evolution in the LLM Landscape | A Conversation with Chen Yu from Yunqi Partners * Google's New Paper Crashes Memory Stocks! 6x KV Cache Compression, Netizens: Silicon Valley Becomes Reality * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native HomeArticlesPodcastsVideosTweets