realworld_go2_task0_task3_smolvla
收藏资源简介:
该数据集是两个子数据集的合并:nullPointerExcept1on/realworld_go2_task0_smolvla(50个episodes,9448帧)和nullPointerExcept1on/realworld_go2_task3_smolvla(25个episodes,5562帧,经过裁剪)。它们对应同一走廊任务,指令为“沿着走廊走”,用于数据规模比较(25 vs 50 vs 75 episodes)。采用SmolVLA格式:仅包含RGB图像,observation.state(move)固定为1.0,action包含vx、vy、vyaw。机器人类型为宇树Go2(unitree_go2)。两个子集的操作方式不同:task0的episodes使用Go2自带的遥控器录制(action=测量身体速度,运动门控录制),task3的episodes通过DualSense手柄经gty-teleop录制(action=速度指令)。每个episode末尾的静止帧已被裁剪。训练时建议使用--rename_map={"observation.images.front": "observation.images.camera1"}。总共有75个episodes,15010帧(日期:2026-09-13)。
This dataset is a merge of two subsets: nullPointerExcept1on/realworld_go2_task0_smolvla (50 episodes, 9448 frames) and nullPointerExcept1on/realworld_go2_task3_smolvla (25 episodes, 5562 frames, cropped). They correspond to the same corridor task with the instruction walk along the corridor, used for data scale comparison (25 vs 50 vs 75 episodes). It adopts SmolVLA format: only RGB images, observation.state (move) fixed to 1.0, action includes vx, vy, vyaw. The robot type is Unitree Go2 (unitree_go2). The two subsets have different recording methods: task0 episodes are recorded using the Go2s built-in remote controller (action = measured body velocity, motion-gated recording), while task3 episodes are recorded via DualSense controller through gty-teleop (action = velocity commands). Static frames at the end of each episode have been cropped. Training recommendation: use --rename_map={"observation.images.front": "observation.images.camera1"}. Total 75 episodes, 15010 frames (date: 2026-09-13).
数据集概述
基本信息
- 数据集名称:realworld_go2_task0_task3_smolvla
- 数据集地址:https://huggingface.co/datasets/nullPointerExcept1on/realworld_go2_task0_task3_smolvla
- 许可证:apache-2.0
- 任务类别:robotics
- 标签:LeRobot、unitree-go2、quadruped、navigation、rgbd、lerobot
- 创建工具:LeRobot
- 机器人类型:unitree_go2
数据集描述
- Homepage:[More Information Needed]
- Paper:[More Information Needed]
- License:apache-2.0
数据集结构
基本参数
- codebase_version:v3.0
- fps:10
- 总回合数(total_episodes):75
- 总帧数(total_frames):15010
- 总任务数(total_tasks):1
- chunks_size:1000
- data_files_size_in_mb:100
- video_files_size_in_mb:200
- data_path:data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
- video_path:videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
- splits:train: 0:75
特征(features)
observation.images.front
- dtype:video
- shape:[480, 640, 3]
- names:[height, width, channels]
- info:
- is_depth_map:false
- video.height:480
- video.width:640
- video.codec:av1
- video.pix_fmt:yuv420p
- video.fps:10
- video.channels:3
- has_audio:false
- video.g:2
- video.crf:30
- video.preset:12
- video.fast_decode:0
- video.video_backend:pyav
- video.extra_options:{}
observation.state
- dtype:float32
- shape:[1]
- names:[move]
action
- dtype:float32
- shape:[3]
- names:[vx, vy, vyaw]
timestamp
- dtype:float32
- shape:[1]
- names:null
frame_index
- dtype:int64
- shape:null
episode_index
- dtype:int64
- shape:[1]
- names:null
index
- dtype:int64
- shape:[1]
- names:null
task_index
- dtype:int64
- shape:[1]
- names:null
配置(configs)
- config_name:default
- data_files:data//.parquet
引用
BibTeX:
[More Information Needed]
备注
- 本数据集是 nullPointerExcept1on/realworld_go2_task0_smolvla(50 个回合,9,448 帧)与 nullPointerExcept1on/realworld_go2_task3_smolvla(修剪后 25 个回合,5,562 帧)的并集。
- 为同一走廊任务,指令为 "Walk down the hallway.",用于数据规模对比(25 vs 50 vs 75 个回合)。
- 采用 SmolVLA 格式:仅 RGB + observation.state=[move](固定为 1.0)+ action=[vx, vy, vyaw]。
- 两半数据的驱动方式不同:
- task0 回合使用 Go2 自带遥控器录制(action = 实测身体速度,运动门控录制)。
- task3 回合使用 DualSense 通过 gty-teleop 录制(action = 速度命令)。
- 每个回合末尾的静止帧已被修剪。
- 训练时使用参数:--rename_map={"observation.images.front": "observation.images.camera1"}。
- 总计 75 个回合,15,010 帧(2026-09-13)。



