Vision-Language Models Are Joint Denoisers in Robotic Action Generation
Vision-Language Models Are Joint Denoisers in Robotic Action Generation
This research introduces JADE, a framework that enables vision-language models (VLMs) and action experts to jointly refine robotic actions through structured action latents. By allowing the VLM to read noisy actions alongside visual observations and language instructions at every denoising step, JADE couples perception with control directly. The method achieves an average success rate of 12.67% on RoboDojo, outperforming other evaluated methods based on pi_0.5, and improves performance on RoboTwin 2.0 and four real-world bimanual tasks by 20.8 percentage points compared to the baseline.
The paper proposes JADE, a method where vision-language models and action experts jointly refine actions during the denoising process, achieving higher success rates on RoboDojo and improved performance on real-world bimanual tasks compared to baseline methods.
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
12.67%
Reported trials
View original evidence
ves an average success rate of 12.67%, the highest among the evaluated $\pi_{0.5}$-based methods, and is competitive with recent VLA and WAM foundation models pretrained at scale. It also improves performance on RoboTwin 2.0 and raises the average success rate across four real-world bimanual tasks by 20.8 percentage points compared with $\pi_{0.5}$.
Open source S5
- paper: 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.
ves an average success rate of 12.67%, the highest among the evaluated $\pi_{0.5}$-based methods, and is competitive with recent VLA and WAM foundation models pretrained at scale. It also improves performance on RoboTwin 2.0 and raises the average success rate across four real-world bimanual tasks by 20.8 percentage points compared with $\pi_{0.5}$.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org