Robotics data: egocentric video, captured and labeled to standard
Robots learn manipulation from people doing real tasks. SoReliable AI captures first-person video of skilled hand work at home and at work, and labels it to your standard.
Every contributor is trained on the capture rules and tested before recording. Every label is checked against a strict quality bar.
The problem
Bad footage teaches robots bad habits
Manipulation data fails in small ways: hands leave the frame, the camera angle drifts, the frame rate is uneven, a task is cut short, or a face appears in a reflection. Each problem makes a clip less useful or unusable.
Labels fail the same way. If the moment of contact is marked a few frames late, the robot learns the wrong timing.
What we provide
What you get with SoReliable AI
Egocentric capture
Head or chest-mounted video of complete tasks, with both hands and the work area in view.Home and workplace skills
Everyday household tasks, and skilled work such as assembly, sewing and wiring at sites that permit filming.Frame-level annotation
Action segments, contact and release moments, slips and recoveries, and an outcome for every episode.Privacy built in
No faces, screens, documents or voices. Contributors are trained to avoid them, and reviewers reject any clip that shows them.Checked against your spec
Resolution, frame rate, camera angle and episode structure are checked on every clip.
An example
A jar-opening clip that looked fine
A contributor records opening a jar for a manipulation dataset.
Capture review: open a jar
- Resolution
- 1920 × 1080
- Episode
- Complete, hands at rest at the end
- Frame rate
- Variable
- What the checks saw: The video is sharp and the task is complete.
- What the expert caught: The reviewer sees the left hand leaves the frame during the twist and the clip was recorded at a variable frame rate. It is sent back with the rule it broke.
Small capture errors make a clip useless for training. Checking every clip against the spec keeps them out.
Illustrative example.
How it works
We prove capability before production
- 1
Turn your spec into training
We break your capture and labeling guide into short lessons with checks.
- 2
Test every contributor
Collectors submit a sample episode and annotators label a clip with an exact answer key before they work.
- 3
Capture and label
Qualified contributors record and label in batches, with reviewers checking each one against the spec.
- 4
Deliver and improve
You receive clips and labels in your format, with rejected footage and reasons logged.
Deliverables
What you receive
- Egocentric video to your capture spec
- Frame-accurate action and contact labels
- Outcome flags for every episode
- Privacy review on every clip
- Rejection logs with reasons
Your rubrics, data, scores and rationales stay yours.
Questions
Questions teams ask
What is egocentric video data?
Video recorded from the point of view of the person doing a task, usually with a head or chest-mounted camera, so a robot can learn how hands handle objects and tools.
How accurate are the labels?
Annotators are tested on clips with exact answers before they work, and contact moments are checked to within a few frames, as your spec requires.
How do you handle privacy?
Contributors are trained to keep faces, screens, documents and voices out of view, and reviewers reject any clip where they appear.
Can you record at workplaces?
Yes, at sites that give permission, with contributors who are qualified for the work and wear the required protective equipment.
Start with one task
A pilot gives you experts proven on your hardest cases, and scores you can check, before you commit to more.
Related
- AI evaluation expertsSoReliable AI gives AI teams evaluation experts from the field the model works in. Doctors review clinical answers. Engineers review code. Lawyers review contracts.
- Expert data for RLHF and fine-tuningYour model learns from the judgments in its training data. SoReliable AI provides RLHF and fine-tuning data created by experts who know your field.
- LLM evaluationSoReliable AI evaluates large language model outputs with experts from your field, against a rubric built from your own quality bar.
Ready to try SoReliable?
Tell us what you are building. Any team, any project, any size. We find the specialists, train them on your standard, and evaluate the work against what your project actually needs.