Data Preparation

Wall-OSS-0.5 training uses LeRobot v3 format data. X Square Robot data collection provides two entry points:

1. Collection via the data collection platform;

2. Collection via the SDK.

For general recording, export, and quality check workflows, see the Data Collection Guide. This page focuses on the LeRobotDataset v3 conversion and data contract required for Wall-OSS training; both collection paths should provide independent conversion scripts to convert raw data into the same format.

Data contract

Before training, verify at least the following:

• Camera keys, order, resolution, FPS, and video encoding;

• Shape, dtype, units, and field order of observation.state;

• Shape, dtype, units, action mode, and field order of action;

• Episode boundaries, task text, timestamps, and indices;

• Missing frames, duplicate frames, loss of synchronization, and abnormal actions;

• Data consistency with key_mappings in the model configuration.

LeRobotDataset v3 typically contains:

• Parquet: Low-dimensional time-series data such as joints, states, actions, and timestamps;

• MP4: Multi-camera videos;

• meta/info.json: Features, shapes, dtypes, FPS, and file templates;

• meta/stats.json: Normalization statistics;

• meta/tasks.jsonl: Task text and task IDs;

• Episode metadata: Length, tasks, and offsets in shared files.

Path 1: Collection via the data collection platform

The data collection platform should handle task creation, device and operator selection, episode recording, result verification, and data export.

For detailed steps, please refer to the Data Collection Guide.

Path 2: Collection via SDK

SDK collection is suitable for developers customizing collection programs, sensor combinations, or task workflows.

For detailed steps, see Data Collection Example.