代码仓库
模型 / 科研
Anchor Align · dwipddalal
Anchor-Align研究VLA微调与表示锚定和语言-动作对齐。
概述
Anchor-Align studies VLA fine-tuning with representation anchoring and language-action alignment.<br><br>Publisher summary：“Anchor-Align (arXiv:2607.13429)：VLA finetuning that prevents behavior cloning from erasing pretrained VLM representations (catastrophic forgetting) and aligns language with actions. OOD generalization on a…”<br><br>Use its repository to inspect the implementation, documentation and supported workflows. This directory record identifies the maintainer, code license and related resources; it does not claim that the project has been installed or tested on your robot.
项目信息
- 维护者
- dwipddalal
- 主要语言
- Python
- 存储库代码许可证
- MIT
- 模型重量许可证
- 未记录
- 数据集许可
- 未记录
- 硬件设计许可证
- 未记录
- 维护
- 未存档在最后记录的元数据中
- 最后观察到的存储库推送
- 2026-08-28
代码许可不会自动涵盖模型权重、数据集或硬件设计。未记录的许可证详细信息需要与发布者确认。
文档部分
- Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment
- Table of Contents
- Motivation
- Failure mode 1 — the pretrained VLM's semantics are silently erased
- Failure mode 2 — the model's language and actions disagree on the same observation
- Method：Anchor-Align
