pbelevich/lego42176_garage_parking

LEGO 42176车库停车数据集是一个用于模仿学习的机器人数据集，专门用于驾驶LEGO Technic 42176汽车进入车库。数据集使用LeRobot v3.0格式记录，通过人类操作员通过蓝牙键盘驾驶汽车，同时IP摄像头记录场景。策略学习目标是将摄像头观察（图像）映射到速度/转向命令（动作）。数据集包含100个episodes，总帧数12,037，持续13.4分钟，帧率为15 FPS，图像分辨率为640x480。机器人类型为lego42176，核心任务是驾驶LEGO汽车进入车库。特征包括观察图像（视频格式，形状[480, 640, 3]）、观察状态（速度/转向归一化值，形状[2]）和动作（命令速度/转向归一化值，形状[2]）。动作空间为连续值在[-0.8, 0.8]范围内，速度正值为前进、负值为后退，转向正值为右转、负值为左转。记录过程涉及LEGO 42176汽车通过蓝牙低能耗连接、IP摄像头通过RTSP流传输、人类从起始位置驾驶汽车进入车库，每个episode代表一次停车尝试。数据集可用于训练ACT策略，硬件包括LEGO Technic 42176汽车、Dahua IP摄像头和MacBook Pro M1 Pro计算设备。The LEGO 42176 Garage Parking Dataset is an imitation learning dataset for driving a LEGO Technic 42176 car into a garage. Recorded with LeRobot v3.0 format, a human operator drives the car via keyboard over Bluetooth while an IP camera records the scene. The policy learns to map camera observations to speed/steering commands. The dataset consists of 100 episodes, with a total of 12,037 frames, duration of 13.4 minutes, FPS of 15, resolution of 640x480, robot type lego42176, and the task is to drive the LEGO car into the garage. Features include observation.image (video in h264 format, shape [480, 640, 3]), observation.state (float32, shape [2] representing normalized [speed, steering] in [-1, 1]), and action (float32, shape [2] representing commanded [speed, steering] normalized to [-1, 1]). The action space is continuous values in [-0.8, 0.8] (normalized from the LEGO hubs [-80, 80] range), where speed positive indicates forward and negative reverse, and steering positive indicates right and negative left. Recording involved connecting the LEGO 42176 car via BLE, streaming from an IP camera via RTSP, human driving from a starting position into a garage using keyboard, with each episode as one parking attempt. The dataset was used to train an ACT policy on Apple Silicon, and hardware includes LEGO Technic 42176 car, Dahua IP camera, and MacBook Pro M1 Pro compute.