x-square-robot/xplanner-benchmark

XPlanner-Benchmark

XPlanner-Benchmark is the portable release of the X-Planner 1,500-episode evaluation benchmark. It contains synchronized multi-view robot-manipulation videos and the episode-level task, subtask, action, scene, duration, and complexity metadata used by X-Planner.

Contents

• —1,500 episodes
• —3,490 MP4 video references
• —167 source dataset identifiers
• —525 unique task names
• —31 task classes
• —41 inferred atomic action labels
• —22.70 total hours of episode video duration (sum of manifest durations)

The release is organized as media/<episode-id>/<view>.mp4. The canonical metadata is in metadata/manifest.jsonl; every media reference in that file is a repository-relative path. metadata/episode_index.jsonl is a compact path index, and metadata/summary.json contains the release statistics.

Dataset preview

The Dataset Viewer shows one row per episode in the test split (1,500 rows). The face_view, left_wrist_view, right_wrist_view, global_view, and side_view columns contain playable videos. A missing camera view is shown as an empty cell. Each row also includes the instruction, task, source dataset, duration, and target subtasks.

data/episodes.parquet is the typed preview table. Its video references are pinned to the verified media revision; the canonical full annotations remain in metadata/manifest.jsonl.

Episode record

Each manifest row contains a stable id, the original source uid, source dataset and group, task instruction, task name, subtask descriptions, semantic attributes, and a camera_videos list. Each camera entry includes the logical view, original camera name when available, relative MP4 path, duration, and byte size. No machine-specific absolute paths are included in the published metadata.

The benchmark has three camera views for 971 episodes, two views for 48 episodes, and one view for 481 episodes. Camera names are source-dependent; use view together with raw_camera when building a loader.

Selection and provenance

The benchmark was selected deterministically from 1,654 eligible candidates using the seed benchmark3-1500-20260818, retaining 1,500 episodes after removing 154 high-redundancy combinations. The source mix is 861 episodes from x2_numeric and 639 from public_named. The selection inventory and source counts are available in the metadata files.

Several source datasets contribute to this combined release, including public robot-manipulation collections and the RoboChallenge Table30 task archives. The exact source identifiers and episode counts are listed in metadata/source_ledger.json and metadata/source_datasets.tsv.

Licensing and responsible use

This is a combined benchmark assembled from multiple upstream datasets. It does not have one universal license. Each source dataset retains its own license, attribution, privacy, and redistribution requirements. Before using or redistributing any subset, consult the upstream terms for every source represented in that subset. The release metadata intentionally records provenance so that this review is possible.

The X-Planner repository code is MIT-licensed; that license does not grant rights to the benchmark media.

Intended use

The benchmark is intended for research on long-horizon robot task planning, event-structured prediction, video-conditioned planning, and whole-episode evaluation. It is suitable for offline analysis and evaluation. It is not a robotics safety certification and should not be used as the sole evidence for deploying a physical system.

Citation

Please cite the X-Planner paper (arXiv:2609.25187) and cite the original upstream datasets used by the episodes in your experiment.

@article{xplanner2026event,
  title   = {X-Planner: Event-Structured Task Planning for Embodied Intelligence},
  author  = {{X Square Robot Team}},
  journal = {arXiv preprint arXiv:2609.25187},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.25187}
}

Version

Release: XPlanner-Benchmark v1.0 (2026-09-16).