静态缓存页面 · 查看动态版本 · 登录
智柴网 登录 | 注册
← 返回话题
✨步子哥 @steper · 2026-05-20 03:28

WorldString Poster

Physical World Models

WorldString

Kunqi Xu, Jitao Li, Jianglong Ye et al. • 2026 • Tsinghua / CalTech / NVIDIA

warning The Missing Link
Video Gen Pixels in → Pixels out
Identity ✗ Dynamics ✗
VS
Scene Recon Geometry → Geometry
Dynamics ✗ States ✗

hub WorldString: Object as State Manifold

Treats objects as continuous state spaces (manifolds) learned from observation.

input
Encoder
Obs → Latent Code
gradient
Latent Space
(State Manifold / DNA)
view_in_ar
Decoder
Code → Geometry

psychology Core Properties
timeline State Manifold
Latent space encodes all valid configurations. Interpolation = Physical motion.
difference Diff. Twin
Fully differentiable structure enables gradient-based policy learning.
touch_app Actionable
Prescriptive, not just descriptive. Encodes "how to change" the object.
healing Robustness
Auto-completes occluded geometry & fills sensory gaps (Material Completion).

swap_horiz Paradigm Shift
Before
Static Snapshot + Separate Dynamics
After
State Manifold (Integrated Dynamics)

analytics Experimental Insights
  • check_circle

    Interpretable Tokens

    Latent queries specialize in object parts (e.g., thumb, palm) consistently across poses.

  • check_circle

    Real-World Soft Bodies

    Successfully models high-DoF non-linear manifolds (Rope, Cloth) preserving volume consistency.

arXiv: 2605.18743 cs.AI

👍 1