Self-Evolution
Interaction outcomes feed back into better world-policy rollouts.
Overview
Worlds that can be seen, touched, and changed.
EvoPhys-World treats a scene as a living system:
it remembers what is there, imagines how actions reshape it, compares possible futures,
and feeds each interaction back into the next rollout.
Capability Shift
From seeing to shaping
Controllable Interaction
Physical response and action branches turn one state into multiple futures.
Human-Centric World Twin
Egocentric interaction data and scene memory build a coherent world twin.
Dimensional Evolution
From 3D worlds to 5D control
Controllable World
Action, causality, and value shape possible futures.
Temporal World
Time connects past, present, and future state.
Spatial World
Scene geometry, objects, layout, and space.
Model Explorer
One base model, two forms.
Demos
From Simulated Worlds to Embodied Actions.
pose
first frame
video
pose
memory pool
video
Physical Interaction
Triangle Roaming
Loop-Return Roaming
Retargeting
Action Rendering
Evaluation Results
Benchmark
| Model Name | WorldScore-Static | WorldScore-Dynamic | Camera Control | Object Control | Content Alignment | 3D Consistency | Photometric Consistency | Style Consistency | Subjective Quality | Motion Accuracy | Motion Magnitude | Motion Smoothness |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| EvoPhys-World 🥇 | 83.45 | 73.78 | 94.56 | 85.52 | 83.37 | 86.78 | 85.56 | 83.24 | 65.12 | 62.47 | 25.18 | 66.00 |
| WorldScape-0.2 🥈 | 81.37 | 73.62 | 75.47 | 87.34 | 73.91 | 87.59 | 89.64 | 90.88 | 64.79 | 60.84 | 24.18 | 81.51 |
| EonWorld 🥉 | 81.08 | 73.37 | 79.51 | 61.00 | 82.12 | 92.76 | 89.18 | 96.70 | 66.28 | 65.35 | 24.75 | 76.81 |
| WorldScape-0.1 | 80.76 | 73.18 | 74.90 | 86.93 | 71.49 | 87.65 | 89.51 | 90.63 | 64.20 | 60.84 | 24.18 | 81.51 |
| FantasyWorld-1.0 | 80.45 | 71.39 | 81.45 | 87.90 | 66.94 | 84.62 | 94.07 | 86.69 | 61.46 | 50.30 | 24.61 | 75.81 |
Worldline Reasoning
Multiple futures
from one world state.
Given different actions, EvoPhys-World imagines different physically grounded futures and selects actionable paths.
Shared state
One scene memory becomes the root of every rollout.
Action branches
Different contact choices bend the same world forward.
Policy value
Imagined futures are compared before execution.
s0
world state
Action A
Action B
Action C
Stable contact
Motion preserves support and advances the task.
Object slip
Friction and angle change the predicted outcome.
Recoverable path
The policy selects a future with better value.
Cup worldline rollout
Team
EvoPhys Team
Do challenging, meaningful things together — to bring Physical AGI into the world.