DIFFRACTION
Human observation · Physical AI · Applied community
Real tasks.
Inspectable training data.
Diffraction builds human-demonstration datasets for robotics and physical AI. We connect task-specific capture, data processing and an applied AI community—with public samples you can inspect before starting a conversation.
Start with the evidence
Two task domains. Different signals. Each release includes its own provenance, quality notes and limitations.
Know what you are using. These are small human-observation samples, not robot-control benchmarks. Estimated annotations are identified separately; neither release claims independently measured annotation accuracy or robot-policy gains. Maintenance has no depth or IMU. Consult each dataset card for available channels, missing data, grouping and license terms.
From a sample to your specification
Bring the task, environments, viewpoints and signals your project needs. We scope capture, processing, quality checks and delivery around an agreed brief. Controlled and distributed capture are configured to the project; robot-arm collection is in pilot buildout.
Better data starts with a better brief.
Define tasks, measured and estimated signals, coverage and acceptance criteria with our practical buyer’s guide.
Read the guide & get the free template →People are part of the pipeline
The Operators Lab brings together builders, engineers, capture operators and physical-AI practitioners. Explore the community to learn, contribute and find project fit. Participation does not guarantee paid work.
Our wider work includes creator content licensing through StreamGenie and business data licensing.
Follow this organization for new releases.
Website · Sample library · Operators Lab · business@diffraction.agency