Large-scale demonstration data
Collection protocols, teleoperation rigs, and quality metrics for demonstration datasets that stay useful past the paper deadline.
Data foundations
for robot learning
Benchmarks go stale, demonstrations don't transfer, and every lab rebuilds the same pipeline. We work on the layer underneath: the datasets, evaluation, and tooling that make robot learning reproducible.
What we work on
Five problems, one layer
Robot learning keeps rediscovering that the bottleneck is data, not architecture. These are the pieces we think are worth building in the open.
Sim-to-real transfer
Measuring what actually survives the reality gap, and building simulators whose failures predict real hardware failures.
Evaluation and benchmarks
Standing evaluation suites with real robots in the loop, so a number on a leaderboard means something on a table.
Manipulation policies
Policy architectures for contact-rich manipulation, studied through the lens of what data they need rather than what they score.
Open tooling
Dataset formats, loaders, and replay infrastructure. Boring software that everyone needs and nobody wants to maintain alone.
2.4M
Demonstrations released
18
Robot platforms
40+
Labs using our data
CC-BY
Default license
Open by default Everything we make
Everything we make
ships with the data
A result you cannot rerun is a press release. Every paper we publish comes with the dataset, the loader, and the evaluation script that produced the number.
- Permissive licenses, no registration wall
- Versioned releases with changelogs and checksums
- Hardware-in-the-loop evaluation anyone can submit to
Releases
Datasets and benchmarks
Built for the boring parts: consistent schemas, honest failure cases, and enough metadata to know what you are training on.
RDL-Manip
Contact-rich tabletop manipulation across 12 embodiments, with synchronized wrist and third-person video.
RDL-Nav
Indoor navigation trajectories collected in occupied buildings, including the failure cases most datasets filter out.
RDL-Bench
A standing evaluation suite with hardware in the loop. Submit a policy, get a real-robot score.
Get in touch
Working on the same problems?
We collaborate with labs and teams collecting robot data in the wild. If you have a platform, a dataset, or a benchmark you want to open up, we would like to hear about it.