G1_Dex3_PickAndPlaceTrash
收藏资源简介:
本数据集包含使用Unitree G1人形机器人搭配Dex3-1灵巧手进行遥操作演示的92个片段,总计198,846帧,约66分钟,采样频率为50Hz。任务为机器人从桌子上捡起一个塑料瓶,走到地上的篮子旁并将其丢入篮中。数据集包含以下字段:观察部分包括头戴摄像头图像(480x640x3,H.264编码,50fps)、43维关节位置(包括双腿、腰部、双臂和双手的关节角度)、14维末端执行器状态(双手位置和四元数)、4维基础方向四元数、3维基础坐标系下的投影重力向量;动作部分包括64维GEAR-SONIC编码器输出的运动令牌(需匹配对应解码器重放)、43维全身控制器产生的关节目标、7维右手关节命令(抓取手)、7维左手关节命令(均为零,左手未使用)。数据集通过VR设备在真实Unitree G1硬件上遥操作采集,使用NVIDIA GEAR-SONIC作为全身控制器,操作者直接控制上半身,控制器处理平衡和行走。适用于人形机器人运动-操作任务的视觉-语言-动作模型后训练,尤其是接受43维人形机器人状态输入的策略。数据集遵循Apache 2.0许可。
This dataset contains 92 episodes of teleoperation demonstrations using the Unitree G1 humanoid robot with Dex3-1 dexterous hands, totaling 198,846 frames, approximately 66 minutes, sampled at 50Hz. The task is for the robot to pick up a plastic bottle from a table, walk to a basket on the ground, and drop it into the basket. The dataset includes the following fields: observation part includes head-mounted camera images (480x640x3, H.264 encoded, 50fps), 43-dimensional joint positions (including joint angles of legs, waist, arms, and hands), 14-dimensional end-effector states (hand positions and quaternions), 4-dimensional base orientation quaternion, 3-dimensional projected gravity vector in base frame; action part includes 64-dimensional motion tokens output by GEAR-SONIC encoder (requires matching decoder for replay), 43-dimensional joint targets from whole-body controller, 7-dimensional right-hand joint commands (grasping hand), 7-dimensional left-hand joint commands (all zeros, left hand not used). The dataset was collected via VR teleoperation on real Unitree G1 hardware, using NVIDIA GEAR-SONIC as the whole-body controller, with the operator directly controlling the upper body and the controller handling balance and walking. It is suitable for post-training of vision-language-action models for humanoid robot locomotion-manipulation tasks, especially policies that accept 43-dimensional humanoid robot state input. The dataset is licensed under Apache 2.0.




