phail-anon/phail-v1.0

PhAIL是一个用于视觉语言动作（VLA）策略的真实机器人评估基准。它包含约950个机器人回合的数据，这些数据在Franka Research 3机器人上使用Robotiq 2F-85夹爪执行bin-to-bin抓放任务，涉及四种对象类型：电池、剪刀、毛巾和木勺。数据集提供同步的外部RGB视频和手腕RGB视频、末端执行器和夹爪的遥测数据，以及每个回合的事件注释。这些回合由四个公开发布的VLA策略（使用各自的默认开源微调配方）和人工遥操作参考生成，还包括用于VLA配方的训练语料库。PhAIL is a real-world robotic evaluation benchmark for Visual Language Action (VLA) policies. It contains data from approximately 950 robotic episodes, where bin-to-bin pick-and-place tasks are performed on a Franka Research 3 robot equipped with a Robotiq 2F-85 gripper, involving four object categories: batteries, scissors, towels, and wooden spoons. The dataset provides synchronized external RGB videos and wrist-mounted RGB videos, telemetry data of the end-effector and gripper, as well as event annotations for each episode. These episodes are generated by four publicly released VLA policies (using their respective default open-source fine-tuning recipes) and human teleoperation references, and also include training corpora for VLA recipes.