RB-VLA adds a persistent belief state to vision-language-action models for long-horizon robot tasks
A new arXiv paper proposes a belief-centric architecture that tracks task progress internally, reporting higher success rates and lower latency than prior vision-language-action models.
In plain words
Robots that follow spoken or written instructions often lose track of what they have already done, so they repeat actions or respond slowly. The RB-VLA paper proposes keeping an internal summary of task history and progress, updated as the robot acts, instead of repeatedly consulting a large vision-language model. The authors report this improves success on multi-stage tasks and speeds up inference.
What was announced
The paper argues that most existing vision-language-action models are observation-driven, relying on short context windows or repeated queries to vision-language models, which the authors link to lost task progress, repeated actions under perceptual aliasing, and high inference latency. RB-VLA instead maintains a compact latent state encoding task-relevant history, dynamics, and object interactions. A vision-language model is queried once per task to supply high-level intent, while the belief state tracks progress and, together with that intent, conditions a diffusion policy for closed-loop control under partial observability, without storing raw observations or growing memory over time.
On the reported benchmarks, RB-VLA achieved 52.5 percent and 37.5 percent higher success rates than pi_0 on multi-stage pick-and-place and stacking tasks respectively, and reduced inference latency by up to five times relative to baselines. Ablations attributed most of the gain to the belief module, with success rising from 32.5 percent without it to 77.5 percent with it. These are the authors' own results and have not been independently verified.
What it does not show
- Evidence is limited to the paper's abstract; the full article was not reviewed, so method details and evaluation setup cannot be checked.
- All performance figures come from the authors' own benchmarks and ablations; no independent confirmation is available.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.