DAVIAN-Robotics/RoboLab-BananaInBowl-30cm-relrate-k1c0-stoch-260711
收藏资源简介:
该数据集名为RoboLab (Isaac Lab) BananaInBowl @30cm,包含从RFCL x RoboLab RL策略收集的rollout数据,已转换为LeRobot v3.0格式,适用于pi0.5(pi05)监督微调(SFT)。数据集记录了机器人操作任务,具体为将香蕉放入碗中的场景,初始随机化距离为30厘米。策略使用relrate k1-c0(速率限制的相对关节动作),采样为随机(stoch)方式。动作空间记录为绝对关节目标(在droid框架中,第7关节偏移为-pi/4),夹爪状态为连续值[0,1](但已被标注为有误,因为实际环境中的夹爪项是二进制的)。成功条件基于K-hold settle,设置success_hold_steps=15,并以成功作为终止条件。数据集包含两个相机视角:over_shoulder_left_camera(外部)和wrist_cam(腕部)。数据集总计5916个episodes,387310帧,帧率为15fps,特征包括动作、观察状态、速度、图像等。注意:该数据集已被取代,建议使用替代数据集RoboLab-BananaInBowl-30cm-relrate-k1c0-stoch-gripbin-260716。
The dataset is named RoboLab (Isaac Lab) BananaInBowl @30cm. It contains rollout data collected from the RFCL x RoboLab RL policy, converted to the LeRobot v3.0 format, and is designed for supervised fine-tuning (SFT) of pi0.5 (pi05). The dataset documents a robotic manipulation task: placing a banana into a bowl, with an initial randomized distance of 30 centimeters. The policy uses relrate k1-c0 (rate-limited relative joint actions), with sampling performed in a stochastic (stoch) manner. The action space is recorded as absolute joint targets (in the DROID framework, the 7th joint has an offset of -pi/4), and the gripper state is a continuous value [0, 1], which has been noted as incorrect since the actual gripper term in the environment is binary. The success condition is based on K-hold settle, with success_hold_steps set to 15, and success is used as the termination condition. The dataset includes two camera perspectives: over_shoulder_left_camera (external) and wrist_cam (wrist-mounted). In total, the dataset contains 5916 episodes and 387310 frames, with a frame rate of 15 fps, and includes features such as actions, observation states, velocities, images, and more. Note: This dataset has been superseded, and the recommended alternative dataset is RoboLab-BananaInBowl-30cm-relrate-k1c0-stoch-gripbin-260716.




