Berkeley-FrodoBots-7K: 7,000 Hours of In-the-Wild Robot Driving Data

A global collection of approximately 7,000 hours of real-world robot data from FrodoBots EarthRover units. The dataset, created by frodobots and last updated in November 2025, is a filtered and cleaned subset of a larger 10k-hour collection, reannotated using the Model-Based Reannotation (MBRA) method to improve action label fidelity.

Use Cases

• Train robot control policies based on the reannotated action labels.
• Benchmark autonomous driving algorithms on in-the-wild data collected globally.
• Develop and test model-based reannotation techniques using the provided methodology.
• Study robot behavior and sensor data across diverse real-world environments.

Strengths

• Approximately 7,000 hours of real-world robot operation data.
• Data has been filtered, cleaned, and reannotated using the MBRA method to improve label fidelity.
• Collected from a fleet of robots operating in diverse global environments.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and specific file formats are unknown, which may limit suitability assessment.

Provenance

Source

frodobots

Collection Method

Collected from a fleet of FrodoBots EarthRover units operating in the wild, then filtered, cleaned, and reannotated.

Time Range

null

Freshness

Last updated 2025-11 16 00:37:57.

Geography

Global (collected around the world).