Downloadable example
How RB Router works
Buy training capability,
not disconnected data.
Robot teams start with a capability gap: a task, an embodiment and a failure mode. RB Router turns that gap into one governed Skill Dataset with a clear path to train and validate.
The operating model
One SkillSpec.
One governed training package.
RB Router turns the robot, task, failure mode and target outcome into the evidence, training rights, commercial terms and access path your team needs to move.
See a Skill Dataset →SKILLROUTER
OutputTraining-ready Skill DatasetRight data mix · one package
Rights & accessAttached before trainingLicense, delivery or hosted access
Evaluation loopPerformance informs the next routeEvidence, not blind retraining
◎
Performance-aware routingUses quality, fit and training evidence to improve the next route
Specify the capability gap
Name the robot, task, failure mode and target outcome — not a folder of files.
Route the evidence
Match human, robot, tactile and synthetic data to the skill and embodiment.
Commit rights & access
Agree training rights, commercial terms and secure delivery before work begins.
Train, measure, learn
Use evaluation evidence to make the next data decision more precise.
The product
A Skill Dataset is
the unit of progress.
It is a governed training package for one capability: the data mix, embodiment fit, training rights, failure coverage and evaluation evidence required to make a training decision.
Illustrative examples: dataset packages, performance signals and pricing are sample routing scenarios, not published customer results or offers.
Skill Dataset anatomy
One unit. A decision-ready training plan.
A Skill Dataset brings together the data, modalities, embodiment fit, license rights, failure coverage and training evidence needed to teach one robot capability.
SKILLConnector
Skill Composer
One skill can be assembled from many providers.
The router assembles human, robot, synthetic and failure data into the best training mix for the target robot and model.
Human tactile12.4K episodes · 94 QA
Robot teleop9.8K episodes · native
Synthetic50K edge cases
Failure / recovery2.8K events
DATA
ROUTER
ROUTER
Generated Skill DatasetConnector Insertion v472.2K episodes · VLA-ready
Example Skill DatasetsFour sample packages across data modalities and robot tasks.
Robot-native
RoboOps Co.
Dual-arm Assembly Teleoperation
Large scale
SimGen
Connector Failure Edge Cases
Expert demo
SkillWorks
Industrial Fastener Expert Skills
Find the evidence that can move a robot skill forward. If the right Skill Dataset is missing, commission the exact data slice through a structured request.
Request a missing skill →When a Skill Dataset is missing
Commission the missing
data slice — not a new pipeline.
RB Router turns the capability gap into a scoped collection brief, routes it to qualified providers and assembles new human, robot and synthetic supply into one training-ready package.
Buyer request #582320K door connector installsTesla-like geometry · RGB-D + F/T + tactile · commercial training rights · delivery in 6 weeks.
DATA
REQUEST
ROUTER
REQUEST
ROUTER
Human expert provider8K expert demonstrationsTactile + RGB-D · 4 weeks
Robot data provider6K teleoperation episodesRobot-native · F/T · 5 weeks
Simulation provider50K synthetic edge casesRare failures · 10 days
Train & evaluate
Own the checkpoint.
Keep control of the data.
Train on your own compute, use hosted jobs, or keep valuable data private with secure access. Return evaluation results to make the next Skill Dataset more precise.
Integration workflow
From robot failure to the right training data.
We are not just a static data catalog. The platform can sit in the training loop: define the skill, route the right data mix, run training, evaluate on the robot and feed performance back into the network.
01
RouteCreate a SkillSpec from the robot, model, task and failure mode.02
TrainUse BYOC data streaming or hosted training jobs with metered access.03
EvaluateBenchmark success rate, completion time, safety and intervention rate.04
FeedbackReturn anonymous lift signals so the router gets smarter for the next team.Hosted Training Access · connector_insertion_v4● RUNNING
Licensed episodes consumed48,620 / 72,200
GPU compute8× H100 · 5.8 hr
Skill dataset read1.8 TB · 6 training runs
Pay-per-training meter$3,842Data access + compute + platform fee
Integrate with the Data Router
Keep your training stack.
Connect it to the Skill Dataset.
Use the API to request data, attach licensing, start a training job and return evaluation signals from the systems your team already uses.
What the platform manages
For robot companies
Use one API to route skills, access licensed data, trigger training jobs and submit performance feedback.
1Search by skill, robot embodiment, model family and failure mode.
2Consume routed datasets as MCAP, LeRobot or customer schema.
3Upload evaluation metrics like success rate, completion time and safety improvement.
One contractOne billOne SDK
For data providers
Keep data in your own storage or on hosted access, register manifests, attach license rights and join demand-side routing.
1Register storage, metadata and skill coverage through a provider API.
2Expose training rights, deployment rights and usage policies.
3Get visibility into downloads, hosted usage and verified training lift.
DistributionLicensingMonetization
Developer-friendly API
Route data + train
from skillrouter import Client client = Client(api_key="sk_live_...") route = client.route.create( skill="connector_insertion", robot="dual_arm_v2", model="vla", failure=["false_engagement", "slip"], budget=50000 ) job = client.training.create( route_id=route.id, trainer_image="company/vla-trainer:v4", compute="8xH100" ) checkpoint = job.wait().checkpoint_uri
Return evaluation feedback
job = client.training.create(
route_id=route.id,
trainer_image="company/vla-trainer:v4",
compute="8xH100",
mode="hosted"
)
metrics = client.evaluations.create(
training_job=job.id,
benchmark="connector_suite_v2",
return_credits=True
)
client.feedback.create(
training_job=job.id,
skill="connector_insertion",
before={"success_rate": 0.61},
after={
"success_rate": 0.84,
"avg_completion_sec": 16.2,
"intervention_rate": 0.06,
"collision_rate": 0.01
},
visibility="anonymous"
)
# earn credits for verified feedbackProvider onboarding
provider.datasets.register(
storage="s3://provider-bucket/assembly-v3/",
manifest="manifest.json",
skill="fastener_installation",
formats=["mcap", "lerobot_v3"],
rights={
"evaluation": True,
"commercial_training": True,
"deployment": True
}
)
# data can stay in provider storagePerformance intelligence
Build a compounding
skill advantage.
The Dataset Performance Graph links data sources, robot embodiments, model families and task outcomes, so teams can see what to adjust before the next training run.
Illustrative performance scenario
See how a routed data mix can be evaluated.
Performance feedback can improve routing, quality scoring and pricing for the next request.
Success rate0
Sample efficiency0
Failure rate0
Compatibility0
Example Dataset Performance GraphDataset × Model × Robot × Skill
Human tactile+21pp modeled
Robot teleop+16pp modeled
Synthetic failures+17% efficiency
VLA modelπ-like policy
RobotDual-arm v2
Connector insertion92/100 compatibility
Routed skill mixBaseline
Security, rights & formats
The rights layer serious
robot data needs.
Physical AI data is not just a file problem. The platform standardizes licensing, secure access, usage metering and training-ready formats so buyers and suppliers can transact with confidence.
Standardized Physical AI License
Make rights clear before training starts.
Every skill package can expose evaluation rights, commercial training rights, deployment rights, derived-model rights and private hosted-access policies. That removes repeated legal work and makes pricing more transparent.
Rights surfaceEvaluation, fine-tuning, pretraining, deployment and internal benchmark scopes.
Usage controlsDownload, hosted access or bring-compute-to-data, all with auditable access policies.
Training-ready formatsMCAP, LeRobot v3, video + metadata and customer schemas with QA validation.
Commercial trainingStandard platform license
- Train up to 3 internal model families
- Commercial deployment allowed
- No dataset redistribution
- Derived checkpoints permitted
Private hosted accessBring compute to data
- Supplier keeps raw data private
- Buyer receives checkpoints + logs
- Usage metering and audit trail
- Higher-trust enterprise workflows
What buyers get
A single interface for licensing, access, routing, billing and performance reporting — even when data comes from multiple providers.
One contractOne billOne schema layer
What providers keep
Control of storage location, visibility, license scope, private listings and whether raw data can be downloaded or only used through hosted access.
Private listingsHosted onlyUsage analytics
Quality and provenance
Each dataset can expose capture method, modalities, QA score, embodiment compatibility, safety annotations and verified training lift.
DataQAProvenancePerformance score
MCAPLeRobot v3Parquet + MP4Usage meteringAudit logRights policy
For data providers
Monetize hard-won robot experience — with control.
Bring teleoperation, expert demonstrations, tactile data, simulation or proprietary history. You set availability and training rights; RB Router connects it to teams with a defined skill need and a serious path to train.
Keep control of raw dataDefine training rightsReach qualified demand
Apply as a data provider →
Controlled capture environmentProvider retains data control
Start with the bottleneck
What skill is blocking
your robot?
Tell RB Router the robot, task, failure mode and target outcome. We turn that constraint into one routed Skill Dataset, a licensed training path and evidence of what to improve next.