jaco_play (TsFile)

Apache TsFile version of lerobot/jaco_play, the CLVR lab's JACO robot play dataset.

Overview

This dataset was created using LeRobot. It records diverse free-form play behaviour of a JACO arm across many tasks, with per-frame joint state, action, reward and termination signals, plus two camera streams (image, image_wrist). Language instructions are stored in the mirrored meta/tasks.parquet.

• Robot: JACO (CLVR JACO play dataset)
• Scale: 1,085 episodes, 77,965 frames, 89 tasks, 10 fps
• Split: train
• Homepage: https://github.com/clvrai/clvr_jaco_play_dataset

Schema (TsFile structure)

The TsFile table is named jaco_play.

Role	Columns
Time	Time, INT64 milliseconds
TAG	episode_index, task_index
FIELD	frame_index, sample_index, next_reward, next_done
FIELD	observation_state_0 … observation_state_7, FLOAT
FIELD	action_0 … action_6, FLOAT

The original 8-element observation.state and 7-element action vectors are flattened into scalar FLOAT measurements. The boolean next.done flag is stored as INT64 (0/1). The source index column is retained as sample_index.

Conversion notes

• Time = round(timestamp * 1000) with millisecond precision; time restarts inside each episode, while episode_index and task_index identify the TsFile device.
• The original timestamp field is omitted because it is exactly represented by Time / 1000.
• The camera streams (observation.images.image, observation.images.image_wrist) are video and are NOT included in this repository; they remain in the source videos tree.
• meta/ is mirrored from the source. Aside from the redundant timestamp column and the excluded videos, no source rows or numeric fields are dropped.

Read example

from tsfile import TsFileReader

with TsFileReader("data/jaco_play.tsfile") as reader:
    print(reader.get_all_table_schemas()["jaco_play"])

Source & license

• Original dataset: https://huggingface.co/datasets/lerobot/jaco_play
• Author: lerobot
• Homepage: https://github.com/clvrai/clvr_jaco_play_dataset
• License: cc-by-4.0