--- language: - en size_categories: - n<1K tags: - robotics - robot-failure-detection - video configs: - config_name: default default: true data_files: - split: train path: - metadata.jsonl - videos/**/*.mp4 --- # CoTaP: Kitchen Robot Task Outcomes This collection contains 200 kitchen robot episodes across 10 tasks, with 10 successful and 10 failed episodes per task. Each row corresponds to one episode and links a third-person video and a head-camera video. Videos are sampled at 10 frames per second. ## Contents - `metadata.jsonl` contains the task goal and detection question, success or failure label, failure type, annotated failure onset and confirmation frames, and an explanation. - `videos//` contains the two camera views linked by relative paths in the metadata. - `robot_logs/.parquet` contains the original per-frame robot state (57 values), action (57 values), tactile readings (90 values), timestamps and indices. Metadata links each episode to its robot log using `robot_log_path`. - `annotations/.json` preserves the collection annotations, including notes and initial-frame/subgoal fields. Video paths are relative to this repository. Task wording in these original annotations may differ from the revised task goals in `metadata.jsonl`. - `meta/features.json` documents the feature types, shapes and channel names. `meta/episodes.jsonl` links each episode to its original collection, selected recording metadata, statistics and original task records. `meta/dataset_info.json` summarizes this release. - Failure frame indices are zero-based. Success episodes have null failure fields. ## Robot logs and frame alignment ```python from pathlib import Path import pyarrow.parquet as pq root = Path("path/to/cotap_public_200") episode = dataset[0] log = pq.read_table(root / episode["robot_log_path"]) state = log["observation.state"].to_pylist() action = log["action"].to_pylist() tactile = log["observation.tactile"].to_pylist() ``` Row k has `frame_index=k` and corresponds to frame k in both episode videos, with an episode-relative `timestamp` of k/10 seconds. This is the recorded frame association; sensor topic ages and recording timing diagnostics are retained in `meta/episodes.jsonl`. The original Parquet files are preserved byte for byte. Their `episode_index`, `task_index` and `index` values are local to their source collection and may repeat across collections. Use the public `episode_id` and linked paths to join modalities. Original channel names are retained in `meta/features.json`; physical units and action control semantics require the collection-system specification. The dataset card selects only the videos and episode metadata for the default Hugging Face configuration. Read an individual robot log with PyArrow as shown above.