Brett showcases a fleet of autonomous robots using Helix 2 to sort packages, with self-charging and fleet coordination.
要点 · TL;DR
数据是人形机器人最大的制约因素，而非算力或制造。 Data is the biggest constraint for humanoid robots, not compute or manufacturing.
人形形态支持跨任务的迁移学习，提升整个机器人队伍的表现。 Humanoid form enables transfer learning across tasks, improving fleet performance.
Figure 4 旨在实现类似 iPhone 的飞跃，围绕数据和 Helix 3 设计。 Figure 4 aims for an iPhone-like leap, designed around data and Helix 3.
核心观点 · Key points
人形机器人最大的限制是数据，而不是算力或制造。 The biggest constraint for humanoid robots is data, not compute or manufacturing.
人形形态能够实现跨任务的迁移学习，提升整个机器人队伍的性能。 A humanoid form factor enables transfer learning across diverse tasks, improving performance fleet-wide.
设备端推理对于速度、延迟以及在无网络环境下的运行至关重要。 On-device inference is critical for speed, latency, and operation in network-denied environments.
Figure 4 将是一次重大飞跃，围绕数据和 Helix 3 设计，旨在打造类似 iPhone 的时刻。 Figure 4 will be a major leap, designed around data and Helix 3, aiming for an iPhone-like moment.
人手是基准；达到人手水平的机器人手对于通用机器人至关重要。 The human hand is the benchmark; a robot hand at parity is fundamental for general-purpose robots.
世界是为人类建造的，使类人机器人成为通用机器的理想形态。 The world is built for humans, making humanoids the ideal form for general-purpose machines.
反共识 · Contrarian takes
基于肌腱的手是局部最优；Figure 终止了该计划并转向新设计。 Tendon-based hands are a local maximum; Figure killed the program and moved to new designs.
更多样化的数据，即使与任务无关，也能提高该任务的性能。 More diverse data, even unrelated to a task, improves performance on that task.
Figure 正在将大部分供应链移出中国，预计下个季度几乎零依赖。 Figure is moving most of its supply chain out of China, expecting near-zero exposure next quarter.
机器人通过点头进行视觉交流，而不仅仅是通过无线消息。 Robots communicate visually with head nods, not just via wireless messaging.
Figure 4 是为 Helix 3 设计的，而 Helix 3 是为数据设计的，使数据成为核心设计驱动力。 Figure 4 is designed for Helix 3, which is designed for data, making data the core design driver.
对新任务零样本泛化的目标最早可能在今年实现。 The goal of zero-shot generalization to new tasks could be demonstrated as soon as this year.
本期章节 · Chapters（共 29）
开场与观看机器人Introduction and Watching the Robot
机器人大脑与训练数据Robot Brain and Training Data
手部相机与条码扫描Hand Cameras and Barcode Scanning
处理器升级与推理限制Processor Upgrade and Inference Limits
端侧模型及其优势On-device models and their advantages
迁移学习与统一模型方法Transfer learning and the unified model approach
布雷特问答与Figure 4演进Questions for Brett and Figure 4 evolution
Figure 4设计评审与能力跃升Figure 4 design review and capability jumps
Figure 4设计与进展Figure 4 Design and Progress
数据是最大制约因素Data as the Biggest Constraint
产品质量与发货Product Quality and Shipping
演示反思Reflections on the Demo
开场赞誉与交接Opening praise and handover
手部重新设计问题Question on hand redesign
布雷特解释腱驱动手部故障Brett explains tendon-based hand failure
从错误中学习与新手指生成Learning from mistake and new hand generation
主持人询问必要性与自由度Host asks about necessity and degrees of freedom
布雷特解释高自由度需求Brett explains why high DOF is needed
手部对称性重要性共识Agreement on importance of hand parity
布雷特描述定制手部设计与工程Brett describes custom hand design and engineering effort
主持人插话与布雷特继续复杂性Host interjects and Brett continues on complexity