01
World Twin
Scene-level memory builds dynamic twins that preserve spatial structure and temporal continuity.
Overview
From watching worlds to moving worlds.
EvoPhys-World moves beyond visual generation and camera navigation, enabling action-conditioned physical interaction, long-horizon prediction, and human-centered policy learning.
02
Physical Interaction
Actions drive object responses, contact changes, and future state generation.
03
World Policy
Human action space connects egocentric perception, hand motion, and embodied control.
5D World Model
Space, time, action, causality, and value.
The model does not only predict how the world looks. It imagines how different actions push the same world state into different futures.
Stage 01
3D
Spatial scene structure, geometry, objects, and layout.
Stage 02
4D
Space plus time, motion, continuity, and future state prediction.
Stage 03
5D
Action-conditioned worldlines with memory, causality, feedback, and policy value.
Core Architecture
One base model, two forms.
A unified state-action world model integrates 4D spatiotemporal memory, next-state prediction, next-action prediction, and a self-evolving interaction loop.
Mode 1
Model as World Engine.
World Engine builds dynamic scene-level twins that can be navigated, manipulated, and physically interacted with.
Mode 2
Model as World Policy.
World Policy learns in a human-centered action space, then maps imagined interaction into embodied control.
Demos
Scene-level controllability in action.
Video slots are ready for project demos. Each block is designed for a short, looped clip with a direct task title.
01
Scene Navigation
World twin roaming
02
Object Interaction
Physical response
03
Local Manipulation
Fine-grained control
04
Scene Manipulation
Long-horizon rollout
05
Human Action Prediction
Action chunk decoding
06
Dexterous Retargeting
Human-to-robot transfer
07
Worldline Rollout
Multiple futures
08
Real Robot Control
Embodied execution
Self-Evolution
Data, model, interaction, and policy evolve together.
Virtual interaction generates experience, policy learning turns experience into action, and real-world feedback closes the loop.
Step 01
Data
Human-centered egocentric observations and interaction traces.
Step 02
World Twin
Scene-level memory creates dynamic worlds for interaction.
Step 03
Interaction
Actions produce future states, contact changes, and outcomes.
Step 04
Policy Evolution
Generated experience feeds embodied policy learning.
Worldline Reasoning
Multiple futures from one world state.
Given different actions, EvoPhys-World imagines different physically grounded futures and selects actionable paths.
S0
A1
F1
Stable rotation
S0
A2
F2
Object slip
S0
A3
F3
Contact failure
Team
EvoPhys Team, Peking University
Led by Prof. Shanghang Zhang.