Unify task instructions across all four RoboValue subsets

#3
by wzlllong - opened

This updates the overall task instructions for all 20 real-world and 15 simulation tasks in both One-Shot and Full-Shot subsets, addressing RoboValue issue #1.

All 35 task definitions were checked. Text changes are needed for 20 real-world and 11 simulation task types; the other four simulation instructions already match the target definitions. Across both shot subsets this changes 62 task-table instruction rows and 3,131 episode task-text rows.

For real-world task_index=1, the instruction explicitly specifies five individual key presses in the order 2 → 8 → 2 → 6 → 8, with a full release and return to the waiting position after every press. For task_index=5, it specifies five presses of the blue button, with the same release and return requirements. The other instructions use the definitions from the supplied 2026-10-23 metadata release package.

Only the existing task instruction field in the four meta/tasks.parquet files and the tasks column in their four episode metadata Parquet files change. The task instruction field is stored as task or as the existing unnamed Pandas index __index_level_0__; its name and schema are preserved. Task and episode indices, row counts, schemas, all other columns, videos, action/state trajectories, subtask names/counts/time boundaries, and READMEs are unchanged. The paper-versus-SIA subtask partition question is outside this update; DS judging is not rerun.

The source revision is 458b08ae82e34da4dfc76070ff1056a3d5eab057. Local verification checks all four subsets and all eight Parquet files, including exact task definitions, episode task synchronization, schemas and every non-instruction column, and SHA-256 hashes. A second independent check compares the upload paths with the complete metadata file list at that fixed source revision.

This is a reviewable PR; submission does not directly overwrite main.

wzlllong changed pull request status to merged

Sign up or log in to comment