# STARRY: Spatial-Temporal Action-Centric World Modeling for Robotic Manipulation

Source: [STARRY: Spatial-Temporal Action-Centric World Modeling for Robotic Manipulation](<https://www.sota2.com/research/paper/69f2e21355a3c32af89fc690>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-05T05:00:04.460Z

Authors: Yuxuan Tian, Yurun Jin, Bin Yu, Yukun Shi, Hao Wu, Chi Harold Liu, Kai Chen, Cong Huang

Published: 2026-04-29T16:13:39.000Z

## Abstract

Robotic manipulation requires reasoning about future spatial-temporal interactions and geometric constraints, yet existing Vision-Language-Action (VLA) policies often leave predictive representation weakly coupled with action execution, causing failures in tasks requiring precise spatial-temporal coordination. We propose STARRY, a world-model-enhanced action-generation policy that aligns spatial-temporal prediction and action generation by jointly denoising future spatial-temporal latents and actions through a unified diffusion process. To bridge 2D visual tokens and 3D metric control, STARRY introduces Geometry-Aware Selective Attention Modulation (GASAM), which converts predicted depth and end-effector geometry into token-aligned weights for selective action-attention modulation. On RoboTwin 2.0, STARRY achieves 93.82% / 93.30% average success under Clean and Randomized settings across 50 bimanual tasks. Real-world experiments show that STARRY improves average success from 42.5% to 70.8% compared with $\\pi\_{0.5}$. These results demonstrate the effectiveness of action-centric spatial-temporal world modeling for spatially and temporally demanding robotic manipulation.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Hand Over Vegetables on Hand Over Vegetables Real-world (Stage 1)](<https://www.sota2.com/research/sota/hand-over-vegetables-on-hand-over-vegetables-real-world-stage-1.md>): Hand Over Vegetables; Hand Over Vegetables Real-world (Stage 1); Success Rate: 85.

- [Hand Over Vegetables on Real-world (Stage 2)](<https://www.sota2.com/research/sota/hand-over-vegetables-on-real-world-stage-2.md>): Hand Over Vegetables; Hand Over Vegetables Real-world (Stage 2); Success Rate: 70.

- [Tidy Up Room on Tidy Up Room Real-world (Stage 1)](<https://www.sota2.com/research/sota/tidy-up-room-on-tidy-up-room-real-world-stage-1.md>): Tidy Up Room; Tidy Up Room Real-world (Stage 1); Success Rate: 0.75.

- [Tidy Up Room on Tidy Up Room Real-world (Stage 2)](<https://www.sota2.com/research/sota/tidy-up-room-on-tidy-up-room-real-world-stage-2.md>): Tidy Up Room; Tidy Up Room Real-world (Stage 2); Success Rate: 65.

- [Wash Baby Bottle on Wash Baby Bottle Real-world (Stage 1)](<https://www.sota2.com/research/sota/wash-baby-bottle-on-wash-baby-bottle-real-world-stage-1.md>): Wash Baby Bottle; Wash Baby Bottle Real-world (Stage 1); Success Rate: 70.

- [Wash Baby Bottle on Wash Baby Bottle Real-world (Stage 2)](<https://www.sota2.com/research/sota/wash-baby-bottle-on-wash-baby-bottle-real-world-stage-2.md>): Wash Baby Bottle; Wash Baby Bottle Real-world (Stage 2); Success Rate: 60.

- [Robot Manipulation on RoboTwin Randomized 2.0](<https://www.sota2.com/research/sota/robot-manipulation-on-robotwin-randomized-2-0.md>): Robot Manipulation; RoboTwin Randomized 2.0; Overall Success Rate: not recorded.

- [Robot Manipulation on RoboTwin Clean 2.0](<https://www.sota2.com/research/sota/robot-manipulation-on-robotwin-clean-2-0.md>): Robot Manipulation; RoboTwin Clean 2.0; Success Rate: not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.26848>)

[PDF](<https://arxiv.org/pdf/2604.26848.pdf>)