Matha Labs / Product roadmap
Start with ten first-person datasets about human life activities. Grow to 100, then 1,000, then 10,000 datasets as collection and quality goals are met. Develop useful models for developer APIs and HumanHUD, with contributions through HumanCapture.
Updated October 4, 2026
Dataset milestones are a proposed sequence without committed dates. Robotics is much later, unscheduled research. Scope and timing depend on collection quality, evaluation results, partner requirements, funding, and validation.
Verified starting point
The deployed glasses HUD provides a compact display interface and D-pad navigation. Its current data is mock; activity capture, camera/microphone integrations, and the live companion-data path remain development work.
Open the glasses HUDMatha Labs v8 shipped the public lab and product site with illustrated robot interaction, a published changelog, and a page-view counter. The illustrations demonstrate intended interactions rather than connected hardware.
Read the release historyOctober 2026 / In development
The website includes a working data example, source directory, and configured account flows. Our focus is the first ten datasets from a person’s point of view, around activities such as eating, sleeping, cooking, and going to the gym. HumanHUD activity models, HumanCapture submissions and payouts, and developer APIs are in development or planned.
September 21, 2026
One public robot episode produces 303 curated records, 151 separately generated records, playable workflow videos, and provenance downloads.
Inspect the working exampleSeptember 21, 2026
Google and email/password providers are configured in a new Matha Labs project. Create an account or sign in through the account page; email signup/sign-in passed a temporary-account test.
Open account accessSeptember 21, 2026
Fifteen externally published datasets are searchable by task and research area, with publisher, access, and licensing links.
Browse the directoryDefine human life activities from the first-person view, curate published sources, develop separately identified synthetic examples for gaps, and prepare opt-in collection. Set acceptance and annotation criteria for each dataset, then develop and evaluate suitable models. One dataset does not imply one model.
Plan task-based video and sensor capture, annotation, and review. Proposed base bounty: $5 per accepted annotated datapoint, split into $3 for capture and $2 for permitted self- or peer annotation. Higher bounties would reflect task demands; collection would close when enough examples are accepted.
Develop and evaluate models around the activity datasets, plan API access for developers, and build toward useful assistance through AR glasses. Compatible HumanHUD capture integrations and the account experience remain development work.
Proposed milestones
Start with ten first-person datasets across activities such as eating, sleeping, cooking, and going to the gym. Define each task and prototype HumanCapture’s opt-in capture, permitted annotation, review, and bounty flow. Keep source, synthetic, and participant-contributed examples traceable.
Evidence to review: defined tasks and acceptance criteria, reviewed examples and annotations, a working capture-to-review flow, and verified payouts before launch.
Expand into more everyday tasks once the first collections meet their quality goals. Reuse clear capture instructions, annotation guidance, and review methods; evaluate activity models for HumanHUD and prepare the developer API.
Evidence to review: 100 documented datasets, reproducible collection and review, evaluation results for selected models, and a documented API when ready.
Grow the library across major activity categories and their variations. Use observed data gaps and model evaluations to prioritize new collection rather than treating dataset count as proof of capability.
Evidence to review: 1,000 documented datasets, category coverage, provenance, annotation quality, and model performance across different activity settings.
Work toward 10,000 datasets that support human activity understanding. Develop and evaluate models as the tasks require, sharing useful capabilities between HumanHUD and developer APIs.
Evidence to review: a verified dataset inventory, sustained quality, and measured task coverage. Dataset totals do not establish a model count or a percentage of human activity understood.
After the dataset foundations, revisit the simulator, robot SDK, and desktop arm concepts. A compatible arm could connect human demonstrations, permitted robot commands, and observed outcomes. These are research concepts, not current delivery priorities.
Evidence to review before advancing: a useful dataset application, command permissions, simulator results, and independently validated physical tasks.
Preserve the mobile arm and AI-shed concepts for later study. Compatible docking, charging, tool storage, and selected cleaning or yard tasks would need their own feasibility and hardware validation.
Evidence to review before advancing: practical usefulness, compatible hardware, mobile-task feasibility, docking trials, and charging validation.
Preserve the earlier four-arm backpack studies and revisit them only after the dataset work and simpler systems demonstrate practical utility. Wearable hardware remains much later research.
Progress depends on demonstrated usefulness and independent mechanical and operational validation.
Bring a human activity, model task, or data gap. We’ll discuss what a first-person collection needs to capture and how to evaluate its usefulness.