Collecting Physical AI Training Data in Latin America

Physical AI teams need enormous amounts of real-world human data — and increasingly they are sourcing it beyond their home markets. Here is why Latin America has become an attractive place to capture it, and what it takes to run collection there well.

Why look beyond your home market

Models that control robots in the physical world are only as good as the diversity of behavior they have seen. A dataset recorded entirely in one country, in a handful of similar homes, teaches a narrow slice of reality. Broadening where data is collected adds new environments, new object instances, and new ways of doing familiar tasks — the variety that helps a policy generalize instead of memorize.

Latin America is a strong addition to that mix. It offers access to a wide range of everyday environments and willing participants at competitive cost, in time zones that overlap comfortably with North American teams. For a lab trying to expand a dataset's coverage without ballooning its budget, that combination is hard to ignore.

The catch: operations are the hard part

The difficulty is rarely the idea — it is the execution. Standing up data capture in a new region means recruiting and vetting people, finding and coordinating real recording sites, briefing participants on exact capture instructions, handling scheduling and payments, running quality control, and organizing deliverables. Building all of that from scratch, remotely, is slow and expensive, and a remote playbook often misses how homes, neighborhoods, and service sites actually operate locally.

What a local capture partner handles

A capable partner owns the operational chain end to end, so the lab can focus on models:

• Recruitment. Sourcing and vetting participants who match the required profile.
• Site coordination. Arranging homes, neighborhoods, and service sites where recordings take place.
• Capture to spec. Running sessions against agreed acceptance criteria for resolution, framing, environment, and task completion.
• Quality control. Reviewing recordings and rejecting clips that miss the bar before they reach the pipeline.
• Structured delivery. Handing over organized data — for example MP4 video with timestamps and session details — ready for the team to transform for its own pipeline.

For the mechanics of what makes those recordings usable, see our guides on egocentric data for robotics and human demonstration data for imitation learning.

Why start in Mexico City

Mexico City is a practical entry point for Latin America: a large, varied metropolitan area with the range of homes and environments a diverse dataset needs, and close time-zone alignment with US-based teams. A partner that began operating there builds its process from local context — how different homes and service sites work, and how to coordinate recordings that feel native to the environment — rather than from a remote template. From that base, capture can expand to more participants, sites, and task categories.

How programs typically start

Most engagements begin with a pilot: a small, well-defined batch that proves the data is useful and the process is sound. Once it works, the program scales by adding participants, sites, and task categories in parallel. Exact timelines depend on task complexity, the participant profile, and the recording setup — so the right first step is usually to scope a pilot rather than commit to volume up front.