Datasets:

THULab
/

fmb

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - tsfile
  - format:tsfile
  - time-series
  - timeseries
modality:
  - tabular
  - timeseries
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/fmb.tsfile
size_categories:
  - 100K<n<1M

fmb (TsFile)

Apache TsFile version of lerobot/fmb.

Overview

This dataset is a LeRobot v3.0 dataset of the FMB robot arm at 10 fps.

Robot: unknown (FMB arm)
Scale: 1804 episodes, 338,188 frames, 10 fps
Split: train
Cameras (not uploaded): image_side_1, image_side_2, image_wrist_1, image_wrist_2

Schema (TsFile structure)

The TsFile table is named fmb.

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

Vector columns are flattened into scalar FLOAT measurements (single precision). 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.
Camera video streams 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/fmb.tsfile") as reader:
    print(reader.get_all_table_schemas()["fmb"])

Source & license

Original dataset: https://huggingface.co/datasets/lerobot/fmb
Author: lerobot
License: apache-2.0