Ahaskar04/cola-pushblock-1033

COLA Push-Block数据集是一个用于多智能体协调任务的机器人演示数据集，包含1033个演示样本。该数据集基于MuJoCo仿真环境，模拟了两个4自由度机械臂在低摩擦桌面上协作推动小方块的任务：蓝色机械臂（A）作为推动者，将方块推过绿色中线；橙色机械臂（B）作为拦截者，在方块滑出红色失败区域前将其停在绿色成功区域内。数据集以缓存扁平化版本提供，包含332,990个时间步的元组，每个时间步包括两个视角的RGB图像（256x256像素，uint8格式）和两个机械臂的关节位置命令（4维，float32格式，归一化到[-1, 1]）。数据分为训练、验证和测试集，总大小约131 GB（未压缩）。该数据集专为行为克隆和协调学习研究设计，适用于多智能体模仿学习、视觉语言动作模型（VLA）和机器人操作任务。数据收集通过单操作员远程操作完成，环境设置强调协调性测试，因为桌面摩擦极低，方块滑动速度快，时间窗口紧张。数据集还提供了快速加载方法、基线性能（如COLA方法达到42%成功率）和局限性说明（如无轨迹边界、单操作员收集等）。COLA Push-Block dataset is a robotics demonstration dataset for multi-agent coordination tasks, comprising 1033 demos. It is based on the MuJoCo simulation environment, simulating a cooperative push-block task where two 4-DoF arms on a low-friction table work together: the blue arm (A) acts as the pusher to shove a small block past the green midline, and the orange arm (B) acts as the blocker to stop the block inside the green success zone before it slides off the red fail zone. The dataset is provided in a cached, flattened version with 332,990 timestep tuples, each containing RGB images from two camera views (256x256 pixels, uint8 format) and joint-position commands for both arms (4-dimensional, float32 format, normalized to [-1, 1]). The data is split into train, validation, and test sets, with a total uncompressed size of approximately 131 GB. It is designed for behavioral cloning and coordination learning research, applicable to multi-agent imitation learning, vision-language-action (VLA) models, and robotic manipulation tasks. Data was collected via teleoperation by a single demonstrator, with the environment emphasizing coordination testing due to the low friction and tight timing window. The dataset also includes quick-start loading methods, baseline performance (e.g., the COLA method achieves 42% success rate), and limitations (e.g., no episode boundaries, single-demonstrator collection).