ForeAct: Geometry-Guided Action Learning from World Representations
ForeAct: Geometry-Guided Action Learning from World Representations
This paper presents ForeAct, a method for robotic manipulation that integrates scene geometry and dynamic evolution within a single-pass representation backbone. By repurposing a pretrained video world model and using learnable tokens supervised by next-frame relative depth, ForeAct eliminates the need for iterative diffusion or rectified-flow inference and avoids auxiliary geometry modules at deployment. The approach achieves 96.6% average success on LIBERO, 63.5% on SimplerEnv-Bridge, and a 74.4% success rate across 90 real-world rollouts without additional large-scale pretraining.
ForeAct introduces a geometry-guided action learning method that repurposes a pretrained video world model backbone, using learnable tokens supervised by next-frame relative depth to encode spatial structure without iterative denoising or external geometry modules at deployment.
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
Reported success rate
74.4%
Reported trials
View original evidence
ency, geometry-aware control without depth input or an external geometry backbone. Without additional large-scale pretraining, ForeAct achieves 96.6% average success on LIBERO, 63.5% on SimplerEnv-Bridge, and a 74.4% success rate across 90 real-world rollouts.
Open source S6
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ency, geometry-aware control without depth input or an external geometry backbone. Without additional large-scale pretraining, ForeAct achieves 96.6% average success on LIBERO, 63.5% on SimplerEnv-Bridge, and a 74.4% success rate across 90 real-world rollouts.
Open source S6
uses learnable tokens supervised by next-frame relative depth to shape the shared video representation through bidirectional self-attention. This compact interface turns the spatial knowledge already encoded in the backbone into a future-aware representation for action generation, with little parameter overhead. The lightweight geometry readout is used only during training and removed at deployment, enabling low-lat
Open source S5
arXiv:2611.11500v1 Announce Type: cross Abstract: Robotic manipulation requires representations that capture both scene geometry and its evolution during interaction. Existing approaches have largely treated these two aspects separately. World action models capture such dynamics by coupling action prediction with future video generation, but their reliance on iterative diffusion or rectified-flow inference leads to h
Open source S2
igh computational cost, while their representations lack explicit spatial structure. Explicitly modeling scene geometry typically requires auxiliary modules that increase both parameter count and computational cost. Conversely, geometry-centric policies primarily recover observed spatial structure and often lack explicit supervision for anticipating future geometric changes during interaction. To address these limita
Open source S3
tions, we introduce ForeAct, which repurposes a pretrained video world model as a single-pass representation backbone for low-latency control. ForeAct directly processes the clean latent of the current observation in one forward pass, eliminating future-frame synthesis and iterative denoising while retaining the backbone's pretrained spatiotemporal priors. Rather than introducing a separate geometry backbone, ForeAct
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org