IOI: Decoupling Kinematics and Physics for Interactive World Models
TLDR
IOI decouples kinematics and physics in interactive world models using analytical priors and learned dynamics for accurate simulation.
Reasoning
The paper introduces a novel hybrid approach that combines analytical kinematic priors with learned physical dynamics, addressing spatiotemporal drift in data-driven methods. Its strengths include explicit kinematic guidance and multi-view aggregation, but evaluation is limited to the RoboTwin benchmark, leaving generalizability unverified.
Read-first score
Read-first score 65.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 56.
Field roles
Rank sensitivity
Stability: volatile; rank range: 415.
Keyword Scores
Deep Analysis
Innovations
- Decoupling kinematics and physics in interactive world models
- Explicit kinematic guidance via forward kinematics from action sequences
- Multi-view Kinematic Aggregation and Injection module for geometry-consistent conditioning
- Synergy between analytical simulator and video generator
Methodology
IOI integrates analytical kinematic priors with learned physical dynamics. It computes forward kinematics from action sequences to render synchronized front, side, and top orthographic projections. A Multi-view Kinematic Aggregation and Injection module fuses these geometric cues and injects them into the video generator, conditioning video generation on deterministic trajectories to model stochastic physical interactions.
Key Results
IOI achieves state-of-the-art simulation performance on the RoboTwin benchmark with robust zero-shot out-of-distribution generalization. It serves as a reliable policy evaluator with success rates aligning with ground-truth physics simulators, and policies trained on IOI-synthesized data match those trained on teleoperation demonstrations.