W²-VLA Explore results ↘

World-to-Wrist Task-Conditioned Future Wrist Modeling for Fine-Grained Robot Manipulation

Yuhao Pan1,* Haosong Peng1,* Zhengshen Zhang2 Zhengyang Yan1
Yalun Dai3 Fushuo Huo7 Chujie Wang4 Tianyu Qi5
Xiucheng Wang6 Nan Cheng6 Wenchao Xu1,†

1The Hong Kong University of Science and Technology · 2National University of Singapore

3Nanyang Technological University · 4Wuhan University · 5Sun Yat-sen University

6Xidian University · 7Southeast University

*Equal contribution. · †Corresponding author.

PPaper CCode MW²-VLA Model CW²-CoT DDataset

W²-VLA turns global task understanding into action-proximal foresight by predicting future wrist latents from a compact, task-conditioned interface and wrist history.

98.5%LIBERO average
60.71%RoboTwin Easy
18.21%RoboTwin Hard
80+ Hzreal-time generation

Abstract

Vision-language-action (VLA) models often treat main-view and wrist-view observations as parallel visual inputs, overlooking their distinct roles in robot manipulation. Fine-grained manipulation, however, benefits from anticipating how wrist-local interactions may evolve under the global task context. To address this limitation, we present World-to-Wrist VLA (W²-VLA), a VLA model for fine-grained robot manipulation with task-conditioned future wrist modeling. Given current multi-view observations and an instruction, W²-VLA contextualizes a set of latent modeling tokens as a compact interface between the VLM and the wrist predictor. Conditioned on this interface and wrist history, the wrist predictor forecasts future wrist latents, which are converted into future-aware context for action prediction. In addition, we propose W²-CoT, a synthesis pipeline that produces structured annotations for manipulation progress, physical transition cues, and wrist-local evidence. These structured annotations provide auxiliary supervision to help shape the task-conditioned latent interface. Experiments on LIBERO, RoboTwin 2.0, and real-world manipulation tasks demonstrate improved fine-grained and contact-sensitive manipulation across single-arm and bimanual settings, while maintaining real-time action-generation above 80 Hz.

01

Directional modeling

Global world context is explicitly routed toward wrist-local dynamics.

02

Latent foresight

Future wrist representations capture contact, alignment, and release.

03

Fast inference

A fixed-length latent interface avoids explicit CoT generation at deployment.

Method

A two-branch policy joined by one task-conditioned interface.

Architecture of W squared VLA with a Qwen VLM world branch, V-JEPA wrist branch, predictor, adapter, and DiT action head
W²-VLA overview. Future wrist clips are training-only targets; deployed inference uses current observations, wrist history, and the instruction.
World
①

Task-conditioned interface

Current multi-view observations and language contextualize 16 latent modeling tokens inside Qwen-VL.

Wrist
②

Future wrist predictor

Frozen V-JEPA wrist-history features are queried by task-conditioned states to forecast local future latents.

Action
③

Future-aware control

A lightweight adapter fuses the predicted wrist context with VLM states for flow-matching action generation.

Benchmark results

Strong across single-arm and bimanual manipulation.

Real-world evaluation

Robust progress, even when the world changes.

Condition
Metric
Paper visuals

From task-level plans to wrist-local evidence.

Real-world rollouts for table cleaning, occluded placement, and bimanual plug insertion
Real-world rollouts span long-horizon cleaning, obstacle-aware placement, and contact-rich bimanual insertion.
Attention visualizations over main and wrist views during simulation and real-world manipulation
Latent modeling token attention tracks task-relevant evidence across main and wrist views. CoT text is decoded for visualization only.
Real-world videos

Three tasks across four visual conditions.

View source videos ↗
Citation

BibTeX

@misc{pan2026worldtowristtaskconditionedfuturewrist,
  title        = {World-to-Wrist: Task-Conditioned Future Wrist Modeling for Fine-Grained Robot Manipulation},
  author       = {Yuhao Pan and Haosong Peng and Zhengshen Zhang and Zhengyang Yan and Yalun Dai and Fushuo Huo and Chujie Wang and Tianyu Qi and Xiucheng Wang and Nan Cheng and Wenchao Xu},
  year         = {2026},
  eprint       = {2608.05369},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url          = {https://arxiv.org/abs/2608.05369}
}