Datasets for life, from the human point of view.

Matha Labs is building datasets for human life activities from a first-person, or egocentric, point of view. We want useful training data to reflect the way people see, hear, and experience the activities they do every day.

We’re starting with ten datasets around everyday activities such as eating, sleeping, cooking, and going to the gym. Our roadmap grows from 10 to 100, then 1,000 and ultimately 10,000 datasets, with collection and quality goals defined for each stage. We will develop and evaluate models around these datasets for planned developer APIs and HumanHUD; dataset counts do not imply a matching number of models.

We start with published dataset sources, develop clearly identified synthetic examples for gaps, and plan to augment the library with opt-in egocentric video and sensor data. Our current website includes a curated source directory and an executed data workflow; source material and generated examples remain separately identified.

Our long-term ambition is to categorize 99.9% of human activity. That ambition is unmeasured: defining the activity taxonomy and evaluating coverage remain part of the research.

HumanCapture is our planned app for people to capture and annotate examples for specific datasets. The proposed base bounty is $5 per accepted datapoint with its annotation: $3 for capture and $2 for annotation. Contributors could annotate their own submissions or earn the annotation portion on other submissions they are permitted to access. More demanding tasks may offer higher bounties, and collection would close once a dataset has enough accepted examples. Capture submissions and payouts are not live yet.

We’re building HumanHUD: an application for AR glasses that combines egocentric models, visual intelligence, and audio intelligence to understand what people are doing and help with it. The existing browser experience remains available through humanhud.ai while activity understanding and capture integrations are in development.

Our product direction connects three tiers: the human, local edge AI, and Cloud AI. The human sets the goal, local intelligence interprets nearby activity, and Cloud AI helps plan and coordinate useful assistance. Robotics is a much later research direction. We preserve our desktop arm, mobile robot, and Robotic Backpack concepts while focusing current work on first-person datasets.

Alongside our product work, we offer consulting in AI strategy, evaluation, prototyping, and technical architecture. We define each engagement around the questions and decisions a team needs to work through.