AI News AI资讯 18h ago Updated 1h ago 更新于 1小时前 50

Generalist AI Releases GEN-1.5: A Robot Foundation Model That Learns New Tasks From One 3–12 Second Demo Generalist AI发布GEN-1.5:一种能从3-12秒演示中学习新任务的机器人基础模型

GEN-1.5 is a robot foundation model that learns new physical manipulation tasks from a single 3–12 second demonstration via "physical prompting" — no fine-tuning or gradient updates required One-shot in-context prompting achieved 59% average success across 10 diverse tasks; just 10 gradient steps on 5 minutes of data raised performance to 83% The model's capabilities (one-shot learning, sim-to-real transfer, human-to-robot imitation) emerged organically from 8+ months of continuous pretraining o Generalist AI发布GEN-1.5机器人基础模型,可通过单次3-12秒演示在30秒上下文窗口中学习新物理任务,无需微调或梯度更新 零样本单次演示提示在10个操作任务上平均达到59%成功率,10步梯度更新(5分钟数据)提升至83% "物理提示"机制从8个月连续预训练中涌现,未进行架构修改、元学习循环或辅助目标设计 模型展现组合泛化、零样本仿真到现实迁移、人类到机器人模仿等涌现能力 当前为研究发布,无公开权重、API或商业化产品,需通过直接合作获取

72
Hot 热度
70
Quality 质量
75
Impact 影响力

Analysis 深度分析

TL;DR

  • GEN-1.5 is a robot foundation model that learns new physical manipulation tasks from a single 3–12 second demonstration via "physical prompting" — no fine-tuning or gradient updates required
  • One-shot in-context prompting achieved 59% average success across 10 diverse tasks; just 10 gradient steps on 5 minutes of data raised performance to 83%
  • The model's capabilities (one-shot learning, sim-to-real transfer, human-to-robot imitation) emerged organically from 8+ months of continuous pretraining on physical interaction data, with no explicit architectural changes or auxiliary objectives designed for them
  • Adaptation compute is dramatically reduced: 10 gradient steps move model weights by less than 0.15%, suggesting fine-tuning reconfigures existing knowledge rather than building new representations
  • GEN-1.5 remains a research release with no public weights, API, or pricing — accessible only through direct partnership with Generalist AI

Why It Matters

This represents a significant step toward scalable robot learning, demonstrating that large-scale pretraining on physical interaction data can produce emergent one-shot learning capabilities analogous to in-context learning in language models. For AI practitioners, it validates the scaling hypothesis in robotics and shows that test-time training in extremely low-data regimes can collapse adaptation costs by orders of magnitude, making real-world robot deployment more feasible.

Technical Details

  • Architecture & Inputs: GEN-1.5 is a large multimodal model that ingests video, sensor data, language, and proprioceptive inputs, maintains a 30-second context window, and outputs 100 Hz action trajectories
  • Physical Prompting Mechanism: A sensorimotor demonstration (sensor streams plus action trajectory) is inserted into the context window via a drag-and-drop interface; the model then performs the task immediately with zero gradient steps
  • Emergent Capabilities: One-shot learning, compositional generalization (chaining two demonstrations with bridging motions), zero-shot sim-to-real transfer, and human-to-robot imitation all emerged without explicit training objectives
  • Performance Metrics: 59% success (±10% std) for one-shot prompting; 83% (±9%) after 10 gradient steps on 5 minutes of data; 66.5% with just one gradient step on one minute of data on held-out tasks
  • Pretraining: Over eight months of continuous pretraining on physical interaction data collected from homes, warehouses, and factories — no simulation data was included in pretraining

Industry Insight

  • The emergence of one-shot learning from scale alone — without architectural modifications or meta-learning loops — strongly supports the scaling hypothesis for robotics and suggests that future robot foundation models may achieve similar capabilities through continued pretraining investment
  • The dramatic reduction in adaptation compute (10 steps vs. tens of thousands traditionally) makes test-time training viable for real-world deployment, potentially lowering the barrier for customizing robot policies across diverse tasks and environments
  • The sim-to-real and human-to-robot transfer results could significantly reduce the data collection bottleneck in robotics, though the current research-only availability means practitioners should monitor for future public releases or partnership opportunities

TL;DR

  • Generalist AI发布GEN-1.5机器人基础模型,可通过单次3-12秒演示在30秒上下文窗口中学习新物理任务,无需微调或梯度更新
  • 零样本单次演示提示在10个操作任务上平均达到59%成功率,10步梯度更新(5分钟数据)提升至83%
  • "物理提示"机制从8个月连续预训练中涌现,未进行架构修改、元学习循环或辅助目标设计
  • 模型展现组合泛化、零样本仿真到现实迁移、人类到机器人模仿等涌现能力
  • 当前为研究发布,无公开权重、API或商业化产品,需通过直接合作获取

为什么值得看

GEN-1.5证明了机器人基础模型可通过纯预训练规模涌现类GPT-3的上下文学习能力,为具身智能提供了新的技术路径。其极低的适配计算成本(10步梯度更新仅改变0.15%权重)可能重新定义机器人策略的部署范式。

技术解析

GEN-1.5是多模态大模型,接收视频、传感器、语言和本体感觉输入,维持30秒记忆窗口,以100Hz频率输出动作轨迹。核心机制"物理提示"将传感器运动示例(传感器流+动作轨迹)拖放至上下文窗口,模型立即执行任务而无需任何训练。

性能数据显示:单次演示提示平均成功率59%(±10%标准差),10步梯度更新(5分钟数据/任务,约50个演示)提升至83%(±9%)。极端情况下,单步梯度更新配合1分钟数据即可在保留任务上达到66.5%成功率。

关键涌现能力包括:组合泛化(两个独立演示可链式组合,模型自动生成中间过渡动作)、零样本仿真到现实迁移(纯仿真演示可直接用于真实机器人)、人类到机器人模仿(人用手演示,机器人复现动作)。

行业启示

机器人基础模型正从"任务特定训练"向"预训练+上下文学习"范式转变,GEN-1.5证明了大规模物理交互预训练可涌现类语言模型的提示学习能力,这为具身智能提供了可扩展的技术路线。

极低适配计算成本(10步梯度更新 vs 传统数万次)意味着机器人部署的经济性将大幅提升,测试时训练(test-time training)在极低数据 regime 下的可行性得到验证。

当前无公开权重和API的研究发布策略表明,机器人基础模型竞争已进入数据飞轮和算力门槛阶段,早期布局者可能通过专有数据引擎建立持续优势,行业需关注数据收集基础设施的竞争价值。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Robotics 机器人 Open Source 开源 Research 科学研究 Agent Agent Training 训练