Dual-Stream Memory for Vision-Language-Action Policies
Dual-Stream Memory for Vision-Language-Action Policies
This research proposes a dual-stream memory system for vision-language-action policies, which enhances execution state tracking and visual evidence recovery. The model achieves 49.22% mean success on sixteen RoboMME tasks, outperforming the FrameSamp+Modul baseline. On real-robot tasks, it reaches 62.5% mean success, with significant improvements in repeated action scenarios. The study highlights the importance of combining execution context with selective visual access, while also identifying persistent challenges in identity tracking.
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
number of tasks
16
View original evidence
d action-conditioned feature reconstruction. A task-conditioned retriever retains frames above a fixed relevance threshold; linear multi-scale fusion encodes each frame into sixteen spatial tokens. We then freeze the temporal compressor and jointly optimize retrieval and policy learning. Across sixteen RoboMME tasks, the complete model achieves 49.22% mean success versus 44.51% for the published FrameSamp+Modul refer
Open source S3number of tasks
4
View original evidence
ence, improving fourteen tasks and all four suite means. On four real-robot tasks, mean success reaches 62.5% versus 37.5%, with the largest gains on repeated actions requiring execution-stage discrimination. These results support combining execution context with selective visual access, while failures under cup swaps expose persistent identity tracking as an unresolved requirement.
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
d action-conditioned feature reconstruction. A task-conditioned retriever retains frames above a fixed relevance threshold; linear multi-scale fusion encodes each frame into sixteen spatial tokens. We then freeze the temporal compressor and jointly optimize retrieval and policy learning. Across sixteen RoboMME tasks, the complete model achieves 49.22% mean success versus 44.51% for the published FrameSamp+Modul refer
Open source S3
ence, improving fourteen tasks and all four suite means. On four real-robot tasks, mean success reaches 62.5% versus 37.5%, with the largest gains on repeated actions requiring execution-stage discrimination. These results support combining execution context with selective visual access, while failures under cup swaps expose persistent identity tracking as an unresolved requirement.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org