Question regarding the training data section of the technical report

[XinyuSun]

The data in Section 3 and Section 4 don’t seem to align. For example:

In Section 4, training stage 1 (General Capability Training) uses 4.8M data samples, but Section 3 only describes 1.6M general image-text data. It’s unclear which specific datasets from section 3 account for the remainder.

Stages 2 (Embodied Spatial-Temporal Training) and 3 (CoT Reasoning) combined about 500K data samples, whereas the "Share Robot" data in Section 3 alone for temporal planning amounts to 1M already.

[tanhuajie]

Hi, @XinyuSun. Thank you for pointing out this concern. The discrepancy arises because Section 3 reports the total constructed datasets and their composition, while Section 4 describes the actual samples we used for training, after refined allocation and balancing across different subsets.

For instance, in Stage 1 (General Capability Training), the 4.8M samples are not limited to the 1.6M general image-text data described in Section 3, but also include around 3.2M foundational spatio-temporal data drawn from multiple subsets. Similarly, the numbers in Stages 2 and 3 reflect the actual selected data portions after re-sampling and proportional adjustment, rather than directly summing the raw dataset sizes reported in Section 3.

Hope it help you well.