ReactWorld: Structuring Interaction Responses for Embodied Video Generation
ReactWorld: Structuring Interaction Responses for Embodied Video Generation
This research presents ReactWorld, an embodied video generation model guided by structured interaction anticipation. It predicts scene changes caused by robot actions while ensuring object responses align with task intent and contact dynamics. The model uses an interaction response field predictor to anticipate future motion and a response feedback adapter to update object features based on predicted displacements. ReactWorld achieves an average score of 77.51 on RBench, surpassing PhysisForcing by 3.29 points across all nine categories, and scores 86.98 on the Domain metric and 81.87 on average on EZS-Bench.
The paper introduces ReactWorld, a video generation model for embodied agents that uses structured interaction anticipation to improve consistency between robot actions and object responses, reporting benchmark scores on RBench and EZS-Bench.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
average score on RBench
77.51
View original evidence
oed local dynamics inputs. It also converts predicted displacements into positional biases for temporal attention. These two mechanisms guide object feature updates and information retrieval across frames, respectively. Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories.
Open source S4Domain score on EZS-Bench
86.98
View original evidence
On EZS-Bench, it achieves a Domain score of 86.98 and an average score of 81.87.
Open source S5average score on EZS-Bench
81.87
View original evidence
On EZS-Bench, it achieves a Domain score of 86.98 and an average score of 81.87.
Open source S5improvement over PhysisForcing on RBench
3.29
View original evidence
oed local dynamics inputs. It also converts predicted displacements into positional biases for temporal attention. These two mechanisms guide object feature updates and information retrieval across frames, respectively. Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories.
Open source S4score on RBench (category unspecified)
77.51
View original evidence
oed local dynamics inputs. It also converts predicted displacements into positional biases for temporal attention. These two mechanisms guide object feature updates and information retrieval across frames, respectively. Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories.
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
oed local dynamics inputs. It also converts predicted displacements into positional biases for temporal attention. These two mechanisms guide object feature updates and information retrieval across frames, respectively. Through interaction anticipation and generative feedback, ReactWorld achieves an average score of 77.51 on RBench, outperforming PhysisForcing by 3.29 points with improvements in all nine categories.
Open source S4
On EZS-Bench, it achieves a Domain score of 86.98 and an average score of 81.87.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org