本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/junfanzhu98/status/2038153945219305812
Summary
A comprehensive recap of the first Robotics World Model Reading Club session in San Francisco (March 2026), covering the paradigm shift from Vision-Language-Action (VLA) policies to World Action Models (WAMs). The session diagnosed the core bottlenecks in robotics as a lack of unified 3D representation, missing physics-grounded simulation, and fragmented data ecosystems rather than model scale. NVIDIA’s Gr00t N2 (7B params on Thor hardware) is cited as the current strongest end-to-end WAM example.
機器人世界模型讀書會第一期旨在探討從 VLA 政策到世界動作模型的根本轉變。核心瓶頸在於缺乏統一的顯式3D表示、物理模擬的 sim2real 差距，以及碎片化的數據生態，而非模型規模不足。
Key Points
- VLA maps observation → action directly; WAM learns latent world → future trajectories → controllable actions
- Pixel-based representations are geometry-unaware and redundant; explicit 3D (point clouds, keypoint tracking) is the emerging backbone
- D4RT achieves 300× speedup by jointly encoding depth, motion, and correspondence in a unified latent field
- Physics gap is the primary sim2real barrier — contact dynamics, deformable objects, and friction remain unsolved
- “FastWAM” skips future prediction at inference time while retaining co-training benefits, enabling low-latency control
- Robotics lacks a “data flywheel” — no equivalent of internet-scale scraping exists for physical interaction
Insights
The session reveals a tension that most robotics papers avoid: scaling alone cannot solve robotics because the bottleneck is representational and physical, not parametric. The “bitter lesson” argument is inverted here — hand-crafted pipelines dominate production (Ambi Robotics suction grippers, fixed motion primitives) while learned generalization remains aspirational. The key insight that world models must eventually compress into Small Language Models (SLMs) for edge deployment reframes the entire scaling narrative.
Connections
Raw Excerpt
Reality cannot be scraped like the internet. It must be sensed, interacted with, and simulated.