01
Data Curation
Pull episodes from your fleet, then use embedding search to surface the edge cases that close distribution gaps.
Data InfrastructurePowering Embodied AI
Sentious turns raw robot logs into training-ready data: native ROS 2 and MCAP parsing, zero-copy WebGPU review of every sensor stream, and export straight to LeRobot v3.0.
- 01 / IngestROS 2 · MCAPROS 2 · MCAP, parsed natively
- 02 / RenderWebGPU, zero-copyZero-copy WebGPU playback
- 03 / ExportLeRobot v3.0LeRobot v3.0 datasets
02 / Platform
Curate, Annotate, and Evaluate
An agentic workflow, reinforced with human feedback — one system from the first raw episode to the model you ship.
- 02
Labeling & Annotation
Video, LiDAR, audio and fused sensor streams in one workspace — with label lineage and production-grade quality control built in.
- 03
Validation & Evaluation
Benchmark and stress-test policies against the scenarios that matter, long before they meet the real world.
- 04
Improve & Scale
Turn evaluation insight into measurable gains, then scale what works across every robot, task and dataset.
03 / Pipeline
From raw robot logs to training-ready data.
Every episode moves through one pipeline — parsed natively, rendered on the GPU, labeled with a model in the loop, and exported in the format your policy trains on.
- 01
Raw sensor data
Cameras, LiDAR, IMU, joint states and audio — straight from the robot or the fleet.
→ .mcap · .bag · video
- 02
ROS 2 · MCAP parsing
Topics decoded natively and time-aligned to a single clock. No conversion step, no lossy export.
→ time-aligned topics
- 03
Zero-copy WebGPU
Every stream rendered straight from GPU memory, so multi-sensor episodes play back at full fidelity.
→ GPU buffers, no copy
- 04
Model-in-the-loop annotation
Models propose; expert reviewers confirm, correct and teach. Every label keeps its lineage.
→ labels + provenance
- 05
LeRobot v3.0 export
Episodes packaged in the format your policy trains on — ready for the next run.
→ Parquet + MP4 episodes
04 / Agentic Feedback & Recommendations
Turn expertise into a continuous advantage.
Sentious agents analyze annotation work as it happens, incorporate reviewer feedback, turn outcomes into insight, recommend improvements and suggest labels automatically — with humans always in control.
- Analyze at scaleFind patterns, failure modes and ambiguous examples across your data.
- Recommend next stepsSpecific, actionable suggestions for labels, boundaries and workflow.
- Auto-label with confidenceModel suggestions arrive with confidence, provenance and human control.
- Continuous improvementEvery decision makes the next suggestion better.
Agent recommendations
Split this action: align → grasp
Segment · 00:11.4
Object is likely a mug
Confidence 88%
Add contact event at 00:12.9
Event
- 01Agent analyzesFinds patterns and suggests labels
- 02You decideReview, adjust and give feedback
- 03System learnsDecisions shape future suggestions
- 04Better resultsHigher quality, faster, every cycle