ReMoT
收藏资源简介:
ReMoT数据集是一个用于强化学习与运动对比三元组的数据集,包含多个子数据集(如agibot、scannet、generaldata),支持机器人夹爪和手臂状态识别、相机视角变化识别以及一般视觉推理任务。数据采用JSONL格式,每个条目包含图像路径、对话消息和真实答案。
The ReMoT dataset is a specialized dataset for reinforcement learning and motion contrast triplet tasks. It comprises multiple sub-datasets such as agibot, scannet, and generaldata, supporting tasks including robot gripper and arm state recognition, camera view change recognition, and general visual reasoning tasks. The data is stored in JSONL format, with each entry containing the image path, dialogue messages, and ground-truth answers.
ReMoT 数据集详情
数据集概述
ReMoT(Reinforcement Learning with Motion Contrast Triplets)是一个用于机器人运动对比学习的多模态数据集,相关论文被 CVPR 2026 接收为 Highlight 论文。
数据集下载
- 数据集地址: ReMoT_data on ModelScope
- 模型地址: ReMoT_model on ModelScope
- 论文地址: arXiv:2603.00461
数据格式
所有数据集使用 JSONL 格式存储,每条数据包含以下字段:
| 字段 | 类型 | 描述 |
|---|---|---|
id |
string | 唯一数据标识符 |
images |
list | 图像路径列表(支持多图像输入) |
messages |
list | 对话消息列表 |
solution |
string | 标准答案,格式为 <answer>ABC</answer> |
示例数据
json { "images": [ "/data/guozeyu/DATA/agibot/resize_images_4/anchor/a_grab_release_656686_000150_35883.png", "/data/guozeyu/DATA/agibot/resize_images_4/positive/p_grab_release_656686_000508_35883.png", "/data/guozeyu/DATA/agibot/resize_images_4/negative/n_grab_release_656686_000039_35883.png" ], "messages": [{ "role": "user", "content": "The image showed to you is what the robot seen by its eyes. In the image, the robotic arm on the left is the robots left arm, and the robotic arm on the right is the robots right arm. Focus only on robot arm/gripper motion across the three images. Please select from the following options whether the left gripper is opened or closed from Image 1 to Image 2? A: Opened, B: No movement, C: Closed. Please select from the following options whether the left gripper is opened or closed from Image 1 to Image 3? A: Opened, B: No movement, C: Closed. Please select from the following options whether the left gripper is opened or closed from Image 2 to Image 3? A: Closed, B: No movement, C: Opened. Answer all three questions above in order. Only return the correct option A, B,or C for each of the three questions in order inside <answer></answer>, e.g., <answer>CAB</answer>" }], "solution": "<answer>CAC</answer>", "id": "idx_resize_4_test_0" }
数据集组成
| 数据集 | 任务类型 | 描述 |
|---|---|---|
| agibot | 机器人夹爪和手臂状态识别 | 判断不同帧之间夹爪和手臂的状态变化 |
| scannet | 相机视角变化识别 | 判断不同帧之间的相机视角变化 |
| generaldata | 通用视觉推理 | 包括相对位置(mm)、目标检测(om)、计数(vl)等任务 |
图像路径追溯说明
由于仅开源 JSONL 文件(不含原始图像文件),可根据文件命名约定追溯原始图像。
1. Agibot 数据集
- 示例路径:
agibot/resize_images_4/anchor/a_grab_release_656686_000150_35883.png - 文件结构说明:
a_move_complex: 任务类型(如grab_release、move_complex等)652288: Agibot 数据集中的场景或任务编号000450: 视频帧编号
通过上述字段可将图像追溯回 Agibot 数据集中的特定场景和帧序列。
2. ScanNet 数据集
- 示例路径:
scannet/crops3/scene0086_02_000300_A.png - 文件结构说明:
scene0086_02: ScanNet 数据集中的场景编号crops3: 第三个裁剪版本000300: 帧编号A: 原始帧图像B: 模拟反向相机运动的裁剪图像
通过上述标识符可在原始 ScanNet 数据集中定位对应图像。




