Ingest any format. Auto-QA every episode. Materialize training-ready datasets in minutes.
Stop wrangling data. Start training policies.
# Materialize a training dataset in one call
from traceplane import Client
tp = Client()
dataset = tp.materialize(
task="pick-and-place",
embodiment="humanoid-23dof",
qa_score_min=0.8,
hz=30,
format="lerobot-v2",
)
# Stream directly into your training loop
for batch in dataset.stream(batch_size=64):
loss = policy.train_step(batch)
Every lab writes custom scripts to convert, validate, and load data. Months of engineering that doesn't train a single policy.
"Record more demos" is the industry default. Nobody knows which episodes are bad until training fails. No automated QA exists.
LeRobot, HDF5, rosbag2, Zarr, RLDS — every dataset speaks a different dialect. Converting between them is a full-time job.
Download TBs to local disk. Write a custom dataloader. Resample, normalize, shard. All before a single gradient update.
Upload episodes in any format — LeRobot, HDF5, rosbag2, Zarr, simulation logs, teleop recordings. We normalize everything to a canonical schema.
Every episode is scored automatically: structural checks (FPS, dropped frames, action dims), kinematic validation, and semantic checks (object presence, hand-object interaction).
Subtask segmentation, object annotations, scene context, frame embeddings, and discretized action tokens — computed once at ingest, served instantly at query time.
Slice by task, embodiment, quality score, environment, or object class. Get training-ready datasets materialized as LeRobot Parquet + MP4 — sharded for your GPU count.
The platform knows what each dimension means — 7-DoF left arm, 3-DoF waist, gripper state. Store once from a 23-DoF humanoid, query as 7-DoF arm-only. Cross-embodiment search just works.
Structural, kinematic, and semantic quality gates on every episode. CI-gatable strict mode. "QA-scored by Traceplane" becomes the industry stamp of quality.
Pre-tokenized actions, pre-computed frame embeddings (SigLIP/DINOv2), pre-sharded for DDP/FSDP. Your training pipeline starts immediately — no preprocessing.
Full lineage tracking. "Training run #47 used dataset v12. Run #48 added 200 episodes." Reproduce any training run from any point in time.
Stream Arrow batches directly into PyTorch/JAX. No "download 2TB first." Resample on the fly — stored at 30Hz, served at 10Hz or 50Hz.
Go from query to training-ready dataset in minutes. Common views are pre-computed and cached. Predicate pushdown at TB scale via Delta Lake.
Bring your own data — teleop logs, simulation runs, fleet recordings. We store, QA, annotate, version, and serve it back training-ready. The core Traceplane experience.
Don't have data yet? Access our library of QA-scored, annotated trajectory datasets — ready to fine-tune your policy out of the box.
Programmatic access to everything. Query, materialize, and stream datasets directly into your training loop. pip install traceplane and go.
Ingest synthetic data from Isaac Sim, MuJoCo, or Genesis. Auto-QA catches sim artifacts. Mix real and sim data with provenance tracking.
Sell your datasets or buy from others. Companies opt-in to share data on the platform — Traceplane handles licensing, access control, and billing.
Need help setting up your data pipeline? Our team works with you to design ingest flows, QA policies, and training data strategies tailored to your stack.
Training foundation models for robots? Query millions of episodes by task, embodiment, and quality. Pre-sharded for your 64-GPU cluster. No more data wrangling.
Need 500 high-quality pick-and-place demos for fine-tuning? Query, filter by QA score, materialize as LeRobot v2. Published datasets with reproducible versioning.
Capturing demonstrations at scale? Upload raw episodes, get automated QA + annotation. Your data becomes instantly queryable and licensable.
We're onboarding design partners now. Tell us about your data and we'll get you set up.