jink-ucla/TRACE

TRACE数据集包含100,000个视觉-语言样本，用于训练TRACE模型，该模型旨在将高级指令转换为用于机器人操作的确切空间坐标。数据集特别关注区域引用和对象引用任务，这对于空间能力的预测和机器人放置至关重要。数据集来源于RoboPoint数据集，包含50,000个区域引用样本和50,000个对象引用样本，并带有增强推理的注释。

The TRACE dataset contains 100,000 vision-language samples for training the TRACE model, which is designed to convert high-level instructions into precise spatial coordinates for robotic manipulation. The dataset focuses specifically on region reference and object reference tasks, which are critical for spatial affordance prediction and robotic placement. The dataset is derived from the RoboPoint dataset, consisting of 50,000 region reference samples and 50,000 object reference samples with reasoning-enhanced annotations.