Dyna Robotics Introduces Dyna-2: A World-Action Model Pre-Trained on 1 Million Hours of Human Video Dyna Robotics 发布 Dyna-2：基于百万小时人类视频预训练的世界-动作模型

TL;DR

• Dyna-2 is a world-action model (WAM) pre-trained on over 1 million hours of egocentric human video, demonstrating that human video can substitute for scarce robot action-labelled data
• A clear scaling law was established on human data (R²=0.919 for MSE, R²=0.865 for accuracy@0.5), which successfully transfers to unseen robot tasks in zero-shot evaluation
• Video prediction specifically drives cross-embodiment generalization: adding video-only data while holding action data fixed reduced zero-shot robot MSE from 0.340 to 0.120 without improving human in-domain metrics
• Joint denoising of video and actions outperformed action-only approaches on all 39 tested robot tasks at every data scale, with action-only models showing severe overfitting as data grew
• Dyna-2 achieves 1.55× success rate and 1.12× grade compared to Dyna-1's VLA, with production deployment available only through Dyna Robotics' vendor-operated robot cells, not as open weights

Why It Matters

This research directly addresses the fundamental bottleneck in robot learning: the scarcity of action-labelled data. By demonstrating that massive-scale human video can be leveraged for robot manipulation through world-model pre-training, Dyna-2 provides a viable path toward general-purpose robotic systems without requiring expensive teleoperation datasets. The findings are particularly relevant for AI practitioners working on embodied AI, as they establish empirical scaling laws and show that video prediction objectives serve as a powerful transfer mechanism across embodiments.

Technical Details

• Architecture: Mixture of transformers with separate tokenization for video and action. Video uses causal masking with DiT layers; action uses bidirectional self-attention with a shallower transformer that joins the video stream early for real-time latency. Proprioception feeds directly into the action transformer.
• Training methodology: Flow matching with shared trunk and two separate marginal velocity fields for video and action losses. The action network deliberately excludes noised video latents at inference, keeping the policy reactive rather than generative.
• Scaling experiments: Nested subsets of exactly 1,000, 10,000, 100,000, and 1,000,000 hours with identical source proportions. Fixed 100-hour validation set. Power law fits: MSE = 0.0691·D^-0.0184 (R²=0.919), accuracy@0.5 = 0.357·D^+0.0203 (R²=0.865).
• Evaluation: Zero-shot testing on 39 tasks across two stationary bimanual YAM platforms (12 internal, 27 from xdof ABC). Post-training on 14 tasks with at most 10 hours of robot data across three embodiments: 6-DOF YAM arms, WUJI-2 20-DOF dexterous hands, and a semi-humanoid prototype.
• Distillation: Video sampling reduced from 10,203 ms to 110 ms on one H100 GPU, making real-time deployment feasible.

Industry Insight

• The vendor-operated deployment model (no public weights or API) positions Dyna Robotics as a robotics-as-a-service provider targeting mid-market service operators in hospitality, commercial laundry, food service, and light assembly—companies with repetitive, stationary manipulation needs across multiple sites.
• The finding that video-only data improves cross-embodiment generalization without affecting in-domain human performance suggests a new axis for scaling robot policies: investing in massive human video datasets may be more impactful than collecting additional action-labelled robot data.
• The threshold behavior observed in Lockbox Key Turning (0% success up to 100k hours, then 90% at 1M hours) indicates potential phase transitions in robot capability that practitioners should account for when planning data collection strategies and deployment timelines.

TL;DR

• Dyna Robotics发布Dyna-2世界动作模型（WAM），在超100万小时第一人称人类视频上预训练，验证了普通人类视频可替代机器人动作标注数据的可行性
• 研究构建了从1千到100万小时的数据阶梯，首次验证了人类数据上的缩放定律可迁移至未见过的机器人数据，且视频预测是迁移的核心驱动力
• 模型架构采用视频与动作分离token化、共享主干的混合Transformer设计，视频使用因果掩码，动作使用双向自注意力，推理时动作网络不依赖预测视频
• 零样本测试在39个机器人任务上表现优异，100万小时数据使"锁盒钥匙转动"任务从0%跃升至90%，远超Dyna-1的1.55倍成功率
• 当前仅以供应商运营系统形式部署（购买Dyna机器人单元），未开放公开权重、API或许可证，目标客户为酒店、餐饮、洗衣等中端服务运营商

为什么值得看

这篇文章首次系统验证了"人类视频缩放定律"向机器人领域的跨形态迁移，为机器人学习突破动作标注数据瓶颈提供了新范式。对AI从业者而言，它揭示了视频预测能力与机器人泛化之间的因果联系，为低成本机器人策略训练指明了方向。

技术解析

架构设计：Dyna-2采用世界动作模型（WAM）架构，基于视频扩散骨干网络，将视频和动作分别tokenize后通过独立的DiT层堆栈处理，两者相互交叉注意力。 proprioception直接输入动作Transformer，视频token使用因果掩码，动作token使用双向自注意力并早期接入视频流，以优化实时延迟。

训练方法：使用flow matching训练，视频损失和动作损失共享主干作为两个独立的边际速度场。动作网络不接收噪声视频潜变量作为参数，确保推理时策略保持反应性而非生成性。联合去噪在所有动作尺度上均优于纯动作训练。

缩放实验：从1千到100万小时构建嵌套子集（保持各来源相同比例），在固定验证集上验证了人类数据上的幂律缩放（R²>0.86），该定律成功迁移至39个零样本机器人任务（R²=0.884），且在1万至10万小时处出现性能拐点。

视频轴贡献：固定5万小时动作数据，仅增加视频预测数据（无动作标签），零样本机器人MSE从0.340降至0.120，而人类域误差未改善，证明视频预训练专门提升跨形态泛化能力。

部署现状：未公开模型权重、API或许可证，仅通过Dyna机器人单元提供供应商运营服务。蒸馏管道将视频采样从10,203ms降至110ms（H100），已部署于酒店、餐厅、洗衣房等生产环境。

行业启示

数据策略转型：机器人学习正从依赖昂贵动作标注数据转向利用海量无标注人类视频，这大幅降低了机器人策略训练的数据门槛，可能加速服务机器人的规模化部署。

架构演进方向：世界动作模型（WAM）通过联合视频-动作预测实现跨形态迁移，优于传统VLA架构，预示未来机器人基础模型将向多模态生成架构演进，视频理解能力成为机器人泛化的关键。

商业化路径：Dyna Robotics选择封闭部署而非开源权重，反映了机器人AI当前更依赖硬件-软件一体化解决方案的商业现实，中端服务市场（ hospitality、餐饮、物流）将成为首批规模化落地场景。