Point Hemis at your logs
MCAP, ROS bags, teleop captures, straight from the bucket your fleet already writes to.
Raw files stay immutable and never leave your cloud.
The problem
Every robotics team rebuilds
the same pipeline. Badly.
Before a single gradient step, sensors must sync, video must transcode, actions must normalize, and all of it must be reproducible.
01
Months to a first training run. Data engineering, not research, sets the schedule.
02
Half of ML time goes to plumbing. Senior researchers babysit ETL scripts instead of training policies.
03
Nothing off the shelf. Ops platforms stop at visualization, exactly where training begins.
The product
Upload your bags.
Get a dataset.
Everything syncs, automatically
Multi-rate sensors align onto one timeline, video is transcoded for random access, action spaces normalize across embodiments.
Embeddings and quality scores computed per episode.
Stream straight into training
Versioned datasets, curated with SQL and vector search, streaming into PyTorch and JAX at full speed.
Every run pins an exact dataset version.
/ Curate
Semantic curation
“All successful bimanual grasps over 30 s” is one query. SQL and vector search over every episode.
/ Version
Versioning & branches
Git-style commits for datasets. Every training run pins an exact version, reproducible by default.
/ Stream
Zero-copy loaders
Arrow batches stream in-process to PyTorch and JAX. No egress hop, no serialization tax.
/ Export
Open at the edges
MCAP, LeRobot and RLDS in and out. Your data stays portable. No lock-in, ever.
Trust
Runs in your cloud.
Hemis workers process logs where they live. Raw files stay immutable and never leave your buckets.
S3
Your buckets stay the source of truth. Hemis reads in place
0
Raw files copied out of your cloud, ever
OSS
Open-source core: converters, format tooling and dataloader
BYOC
Full control plane deployable inside your own cloud
The model
Priced like the pipeline
it replaces.
Open-source core, three simple ways to pay.
Usageper TB transformedMetered transform, indexing and embedding. Starts with your first upload.
Platformper seat / monthSearch, review, splits and dataset diffing for the whole ML team.
Enterpriseannual contractManaged deployment inside your cloud. Your buckets, our control plane.
Design partners
Your fleet's data plane, run for you.
We're onboarding a small group of design partners and hardening the product on their real fleets. Tell us about your robots.
Talk to the founders