170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人

Summary

This LateTalk episode brings Chen Zhe Peter of AlphaEast back for a 2026 Q2 embodied-intelligence review covering humanoid marathon performance, Figure AI logistics sorting, dexterous hands, Cosmos 3, Physical Intelligence’s Pi 0.7, and Generalist Gen 1. The discussion frames Physical AI as a system competition across hardware reliability, robot data, manipulation infrastructure, model route, and early commercial scenes. Its main synthesis is that World Models do not simply replace Vision Language Action Models; productized world models, action models, and VLA policies are likely to converge through World Model VLA Fusion.

Key Claims

• Peter treats Q2 as a move from world-model research language toward product-level releases, with Cosmos 3 presented as a more productized and open omni-world-model example.
• The episode says humanoid marathon results, especially Honor’s autonomous-navigation wins, show that high-end manufacturing, motors, cooling, capital, and organization capacity matter as much as isolated algorithms.
• Humanoid Robot Marathon is compared to F1: it is not a direct sales scene, but it stress-tests heat dissipation, power, reliability, motion control, and navigation.
• Figure AI’s livestream of three robots sorting roughly 130,000 packages over more than 100 hours is treated as a more commercially legible humanoid-robot demo than dancing or running.
• Robot Logistics Sorting is attractive because real package flows include soft bags, fallen objects, odd shapes, labels that need turning or flattening, and other tail cases that standard machine vision plus fixed arms struggle to handle.
• Robot Teleoperation and Remote Takeover is not dismissed as fraud; Peter argues teleoperation is both a data-collection method and a likely industrial supervision pattern, closer to Robotaxi remote takeover than to a hidden demo trick.
• The episode treats Dexterous Manipulation as a new bottleneck and market layer: hand structure, degrees of freedom, drive method, sensing, heat, retargeting, and hardware supply affect what data and policies can be learned.
• 5G Robotics and other dexterous-hand suppliers are framed as possible infrastructure providers if they can make high-DOF hands stable, affordable, and widely available for research.
• The source says full-body control and data collection are shifting from Aloha-style real-robot teleoperation toward UMI-style collection, egocentric video, whole-body motion capture, and dexterous-hand data, extending Embodied Robot Data Paradigms.
• Physical Intelligence’s Pi 0.7 is described as a VLA route that adds a lightweight world-model component, using future image prediction to influence action generation.
• Generalist Gen 1 is presented as deliberately resisting the world-model or VLA label, because it wants to emphasize direct training on physical interaction data.
• Google DeepMind’s ER 1.6 is framed as stronger spatial and task understanding rather than a direct robot-control policy, suggesting a Robotics Android strategy where multiple robot bodies call a shared software brain.
• OpenAI’s restarted robotics effort is treated as a signal that video generation, world simulation, and robot models are becoming more strategically connected.
• The episode argues that robot-body companies, model companies, dexterous-hand suppliers, data companies, and general foundation-model labs are all contesting where the durable value layer will sit.

Key Quotes

“不是简单颠覆 VLA” — Peter’s view of how world models relate to current robot policies.

“零到一展示” — the episode’s framing of Figure’s logistics-sorting livestream.

“不想被贴标签” — the source’s description of Generalist’s resistance to VLA or world-model categorization.

Connections

• LateTalk, Chen Zhe Peter, and AlphaEast — show, recurring guest, and investment context.
• Physical AI, Embodied AI, Humanoid Robot Commercialization, and Humanoid Robot Marathon — industry frame and hardware stress-test theme.
• Figure AI, Xingdong Era, Robot Logistics Sorting, and Robot Teleoperation and Remote Takeover — logistics-sorting and industrial deployment thread.
• Dexterous Manipulation, 5G Robotics, Genesis Robotics, Tesla, and Unitree Robotics — manipulation hardware, data, and ecosystem thread.
• World Models, Cosmos 3, World Action Models, Vision Language Action Models, and World Model VLA Fusion — model-route synthesis.
• Physical Intelligence, Generalist, OpenAI, Google DeepMind, and Nvidia — model-company and foundation-lab competition around robot brains.
• Real Robot Data Strategy, Embodied Data Pyramid, Physical World Data Flywheel, and Embodied Robot Data Paradigms — data-collection and data-recipe implications.
• Production Robot Scenario Selection, Embodied AI Value Chain, and AI Commercialization Pressure — commercial scene selection and value-chain pressure.

Contradictions

• No direct contradiction found. The source mainly extends existing World Models, Embodied AI, Physical AI, and Humanoid Robot Commercialization pages with a Q2 2026 industry snapshot.
• Productive tension to track: Aether AI’s Causal World Models source argues that world-model usefulness depends on causal variables and transition dynamics, while this source emphasizes productized omni-world-models and fusion with VLA-style policies. The routes differ in emphasis but are not mutually exclusive.
• Open issue: Figure AI says its Helix 02 system performed the livestream autonomously, but the episode records public debate about possible teleoperation. The wiki should treat full autonomy as a company claim unless independently corroborated by later sources.