RePlan-VLA adds progress-aware replanning to robot policies
A new arXiv paper proposes a framework that lets vision-language-action models handle long-horizon manipulation tasks without retraining the underlying policy.
In plain words
Vision-language-action models are AI systems that let robots see, understand instructions, and act, but they tend to struggle when a job requires many linked steps. RePlan-VLA is a wrapper that watches the robot's progress and adjusts the plan when outcomes go wrong, while leaving the original model untouched.
What was announced
The authors argue that VLA policies should not be burdened with long-horizon reasoning, memory, action generation, progress tracking, and error recovery all at once. Instead, RePlan-VLA keeps the underlying policy unchanged and uses its short-horizon manipulation skills inside an execution-aware framework. The system explicitly monitors the robot's motion to evaluate progress and adaptively revises the future plan based on observed outcomes, requiring no additional fine-tuning of the policy itself.
In experiments reported by the authors, the framework improved success rates by 21.9% on RoboCasa365 and 10.4% on CALVIN over the next strongest baseline. They also validated the approach on a real Franka Research 3 robot. These figures come from the paper's own abstract and have not been independently verified.
What it does not show
- Evidence is limited to the paper's abstract; full methods, baselines, and experimental details are not available for verification.
- Reported success-rate improvements are the authors' own claims and have not been independently confirmed.
Sources
Source-support review completed 20 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 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.