Skip to content

Build at machine scale.

Keep one copy of your video, sensor streams, agent traces, spatial data, and rows, in your own bucket. Analytics, search, and training read it straight from object storage.

signal_activity > 0.12 → nominal_grasps · tag v41 · 81,006 of 840,000 momentsepisode_7f31c2 · t+3.2s · member · tag v41

Multimodal, multi-rate

System limitations force fragmentation.

Video runs at 30 Hz, LIDAR at 10 Hz, and IMU at 200 Hz. Most systems cannot query all three on their own clocks. Each tool keeps its own copy of the data instead. Agent traces and spatial data break the same way: spans nest inside spans, and points share coordinates with the polygons on a map.

In LeRobot, a SmolVLA run waits 1 s for the data loader for each 0.7 s of backprop.

LeRobot’s change from dataset format v2.0 to v3.0 left datasets on the Hub with no direct conversion path.

Training input · localUntracked

datasets/

failed_grasps_v3/

failed_grasps_v3_rebalanced/

failed_grasps_v3_rebalanced_final/

full export

840k moments

notebooks/

score_failures_v2.ipynb

last run unknown

Checkpoint_18400Dataset definition → missing
Moving the frontier

Past a threshold, speed changes optimal behavior.

Teams copy data out of a Parquet lake into faster systems because the lake is too slow for some jobs. Each result below removes one of those copies.

Analytics

5–60×

Faster scans for analytics and training

Spiral scans data 5× faster than Parquet on CPUs, and up to 60× faster on GPUs.

FastPair · arXiv, Sep 2026

Search

100×

Faster indexes, lookups, and selections

Hot random access on Spiral is 100× faster than on Parquet. Search and lookups read the table itself, so you keep no separate key-value store.

bench.vortex.dev

Training

94.4 Gbps

Training straight from object storage

Spiral reads object storage into GPU memory at 94.4 Gbps on a 100 Gbps link. Training jobs read the bucket and skip the staged copy on local disk.

Spiral blog · Sep 30, 2026

Workflows

Build a learning machine.

Find a failure, extract the signal, curate the fix, and train on it.

Search video, force, and joint streams across every recording at once. The query and its matches are commits on a branch.

Investigate

Find every failure like this one.

An Enrichment is a derived column, computed by a model or an expression for every row of a table. Spiral appends it without rewriting the table and computes it for new rows as they land.

Enrich

Add columns without rewriting the table.

Define a training set as rules over your data, and test it before you tag it. Training and evaluation read it at the tag.

Curate

Curate and test a training set live, and reproduce it forever.

A recipe weights each training set and fixes the seed, so a rerun reads the same batches in the same order. A stopped run resumes with the samples it would have read.

Train

Train on any mix of your data.

Reversibility & isolation

Agents need commits and rollback even more than people do.

Every change in Spiral is a commit. An agent works on its own branch, and a person reads the diff and merges it. A bad change rolls back to the commit before it.

12 weeks, 3 people, 1 agent

  1. slip_distance v3 recomputed corpus-wide

    jun · tracker_v3 · depth_v5 · 840k moments

  2. eval holdout frozen

    priya · no_eval_overlap · pass · tag eval-v4

  3. rebalance: sample near-misses 2×

    maya · maya/rebalance

  4. drop 87 moments where glare read as slip

    maya · maya/rebalance

  5. drop all recovery demos

    scan-agent · closed-loop dipped · discarded · scratch/no-recovery

  6. merge maya/rebalance

    maya · near-misses sampled 2×

  7. tag failed_grasps v13

    maya · +214 −87 vs v12 · checks 12/12 · tag v13

How Spiral works

Spiral is what the lakehouse was supposed to be.

The lakehouse promised one copy of your data for every workload. Most teams got a warehouse for their tables, a lake of loose files for everything else, and a catalog, indexes, and a data loader to build themselves. Spiral is all of it in one system. It runs in every major cloud, and most of our customers already use it in more than one.Use Spiral when more than one system reads the same data, or when you have nested, complex, or multi-rate data streams. Use it when your GPUs wait on I/O, or when agents write to your data. If one Postgres table holds your data and nothing trains on it, you do not need Spiral.