abdul004/molmoact2-so101-zero-shot-eval
收藏资源简介:
该数据集包含在SO-101机器人手臂上使用AllenAI发布的MolmoAct2 SO-100/SO-101检查点进行零样本评估的跟踪记录。它是一个评估工件,而非训练数据集或模型检查点。数据集记录了通过远程异步推理设置(使用LeRobot MolmoAct2集成)运行模型时产生的轨迹,包括视频、GIF预览、元数据和原始事件日志。评估结果显示,尽管基础设施路径正常工作,但策略在行为层面存在失败,例如初始运动后停滞或产生接近当前机器人姿态的目标。数据集主要用于定性检查和调试,适用于机器人学、视觉-语言-动作和零样本评估研究。
This dataset contains evaluation traces from running the AllenAI released MolmoAct2 SO-100/SO-101 checkpoint zero-shot on an SO-101 robot arm using the official LeRobot MolmoAct2 integration plus a remote async inference setup. It is an evaluation artifact, not a training dataset or model checkpoint. The dataset records traces generated during model execution, including videos, GIF previews, metadata, and raw event logs. Evaluation results show that while the infrastructure path worked, the policy exhibited behavioral failures, such as stalling after small initial motion or producing targets close to the current robot pose. The dataset is intended for qualitative inspection and debugging, suitable for research in robotics, vision-language-action, and zero-shot evaluation.




