README.md

---
license: cc-by-sa-4.0
pretty_name: FrodoBots-Mini-4K
task_categories:
  - robotics
  - other
tags:
  - robotics
  - teleoperation
  - autonomous-driving
  - sidewalk-robot
  - earth-rover
  - multimodal
  - video
  - gps
  - imu
size_categories:
  - n>10K
---

# FrodoBots-Mini-4K

**~4,000 hours** of real-world teleoperation data from **Earth Rover Mini / Mini+** sidewalk robots,
driven by a global operator network across **29 countries**. Each ride bundles synchronized **camera video**
(front, and rear when available), **two-way audio**, and time-aligned **GPS**, **IMU**, and
**drive/control (DRV)** streams.

Third public FrodoBots dataset, after
[`BitRobot/FrodoBots-2K`](https://huggingface.co/datasets/BitRobot/FrodoBots-2K) and
[`BitRobot/Berkeley-FrodoBots-7K`](https://huggingface.co/datasets/BitRobot/Berkeley-FrodoBots-7K).
Like the 2K release it ships **raw, unannotated** per-ride folders — packaged here as tar shards hosted
directly on the Hub.

## At a glance

| | |
|---|---|
| Rides | **14,141** |
| Robots | 453 |
| Countries | 29 |
| Date range | 2024-12 → 2026-07 |
| **Hours (teleoperation session time)** | **3,965 h** |
| Front-camera video | 3,231 h |
| Rear-camera video | 1,909 h (73.9% of rides have a rear camera) |
| Size | ~2.0 TB (1,139 tar shards) |

Every ride includes DRV + GPS + IMU + at least one camera with generated per-frame timestamps.

## The robot — Earth Rover Mini / Mini+

A compact, low-cost teleoperated sidewalk rover. Hardware is open:
[github.com/frodobots-org/earth-rover-mini](https://github.com/frodobots-org/earth-rover-mini) ·
[buy one](https://shop.frodobots.com/collections/earth-rovers).

| Subsystem | Spec |
|---|---|
| Compute host | Rockchip **RV1106G3**, 256 MB DDR, 128 MB NAND, Linux |
| Microcontroller | **STM32F407** (RT-Thread RTOS), closed-loop PID motor control |
| Cameras | Front + optional rear, H.264 **HLS**, adaptive up to **1280×720** (commonly 1024×576 / 480×270), **~20 fps**, ~16 s segments |
| Audio | Two-way — robot **microphone** + operator **speaker** |
| IMU | **MPU-6050** (3-axis accelerometer + 3-axis gyroscope) + magnetometer / compass |
| Positioning | Onboard **GNSS** (latitude, longitude; raw stream also carries altitude, HDOP, fix quality, speed) |
| Drivetrain | 4 wheels / 4 motor channels; commanded by **linear + angular velocity** (UART Control Protocol) |
| Connectivity | **4G/LTE** (Quectel modem) + Wi-Fi AP |
| Battery | 18650 cells, 6000 mAh |
| Firmware in this set | HW versions 6.0.x – 6.2.x |

### How the sensors map to the data

| Robot component | File / columns |
|---|---|
| Motor command (linear/angular velocity) + 4 wheel RPMs | `control_data_*.csv` → `linear, angular, rpm_1..4` |
| GNSS | `gps_data_*.csv` → `latitude, longitude` |
| MPU-6050 + magnetometer | `imu_data_*.csv` → `compass, accelerometer, gyroscope` |
| Front / rear camera | `recordings/*_video.*` + `front/rear_camera_timestamps_*.csv` |
| Mic / speaker | `recordings/*_audio.*` + `mic/speaker_audio_timestamps_*.csv` |

## Geographic distribution

Driven by a worldwide operator network across 29 countries. Country is from the robot registry (96% of rides
have a known country; the rest are `Unknown`).

![Rides by country](https://huggingface.co/datasets/BitRobot/FrodoBots-Mini-4K/resolve/main/geographic_distribution.png)

| Country | Rides | Hours |
|---|---|---|
| China | 4,314 | 516 |
| Philippines | 4,180 | 1,344 |
| United States | 1,484 | 818 |
| Vietnam | 1,483 | 759 |
| India | 404 | 127 |
| Kenya | 288 | 44 |
| Mexico | 189 | 34 |
| Taiwan | 185 | 18 |
| New Zealand | 163 | 17 |
| Panama | 113 | 18 |
| Canada | 111 | 27 |
| Costa Rica | 88 | 11 |

…plus Malaysia, UK, Germany, Botswana, Australia, Brazil, Mauritius, Indonesia, Sweden, Turkey, Nigeria,
Chile, Argentina, Estonia, Netherlands, Singapore, Japan. Top cities include Antipolo City (PH), Wuhan (CN),
Kon Tum (VN), Chandler (US), Bengaluru (IN), and Machakos (KE).

## Layout

```
data/
  minirover-0000.tar        # tar shards, <= ~1.9 GB each, each holds many complete ride folders
  ...
  minirover-1138.tar
metadata.parquet            # one row per ride (includes the shard each ride lives in)
```

Each ride folder inside a shard:

```
ride_<rideId>_<botRef>_<UTCstart>/
  control_data_<id>.csv              # DRV: linear, angular, rpm_1..4, timestamp (epoch SECONDS)
  gps_data_<id>.csv                  # latitude, longitude, timestamp (epoch MS); (1000,1000)=no fix
  imu_data_<id>.csv                  # compass, accelerometer, gyroscope (JSON arrays), timestamp (MS)
  front_camera_timestamps_<id>.csv   # frame_id, timestamp (epoch SECONDS)
  rear_camera_timestamps_<id>.csv    # present when a rear stream exists
  mic_audio_timestamps_<id>.csv      # robot microphone
  speaker_audio_timestamps_<id>.csv  # operator speaker
  recordings/                        # HLS: *.ts segments + *.m3u8 playlists (front/rear video, audio)
```

`metadata.parquet` columns: `folder, ride_id, device_ref_id, start_utc, source, db_dur_sec, n_epochs,
hardware_version, drive_mode, country, city, front_footage_sec, rear_footage_sec, has_rear_camera,
has_control, has_gps, has_imu, has_front_ts, has_rear_ts, has_mic_ts, has_speaker_ts, shard`.

## Usage

```python
from huggingface_hub import hf_hub_download
import pandas as pd, tarfile

meta = pd.read_parquet(hf_hub_download("BitRobot/FrodoBots-Mini-4K", "metadata.parquet", repo_type="dataset"))
row = meta.iloc[0]
shard = hf_hub_download("BitRobot/FrodoBots-Mini-4K", row.shard, repo_type="dataset")
with tarfile.open(shard) as t:
    t.extractall("out", members=[m for m in t.getmembers() if m.name.startswith(row.folder)])
# -> out/<ride folder>/ with CSVs + recordings/
```

See `earth_rover_mini_dataset.ipynb` for stats, distributions, and a full "load one ride" example
(aligning a camera frame to its GPS/IMU/DRV readings).

## Notes

- **Hours**: the headline (~4,000 h) counts teleoperation session duration (session end − start). Actual
  recorded front-camera video is 3,231 h (rear adds 1,909 h more); a few sessions were left open and slightly
  inflate raw session time.
- **Recordings** are de-vendored HLS. `.m3u8` playlists carry the wall-clock start (`TIME=`) and per-segment
  durations; per-frame timestamps are in the `*_timestamps_*.csv` files.
- **GPS** retains `(1000,1000)` rows meaning "no fix" — filter these out before use.
- `country` is from the robot registry; robots without an entry are `Unknown`.
- Audio-only sessions (camera never streamed) are excluded.

## License

The public **FrodoBots-Mini-4K** dataset is released under the **CC BY-SA 4.0** license. If you're interested in
additional datasets similar to FrodoBots-Mini-4K, be it for commercial or academic purposes, please contact us.

## Citation

If you use this dataset, please cite:

```bibtex
@misc{frodobots_mini_4k_2026,
  title={FrodoBots-Mini-4K: Earth Rover Mini Teleoperation Dataset},
  author={BitRobot},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/BitRobot/FrodoBots-Mini-4K}}
}
```

## Commercial Access and Custom Data

For additional coverage, commercial rights, evaluation data, or custom data collection, please contact us.

## Contributions

The team at BitRobot created this dataset, including [Michael Cho](https://x.com/micoolcho) and
[Santiago Pravisani](https://x.com/sanatem).