For Researchers
Foundation models and robot policies are bottlenecked by the volume and diversity of data. Every Chameleon experience is a first-person human demonstration captured across a wide range of real tasks and environments, with the tracking and language annotations policy learning needs. Access data from our catalog of experiences, or tell us the skills you need, and our creator network will capture them.
Diverse tasks and environments — Generalization comes from diversity, not repetition. Chameleon captures people performing real tasks across a wide span of skills, settings, and styles, so your models see the long tail of how work actually happens instead of a handful of lab scenes. The task set grows every day as creators publish new experiences and people perform them.
Egocentric capture with 3D head and hand tracking — Every experience is recorded from the performer's point of view, the same first-person frame your policies act in. Dense annotations come with it: camera poses, 3D head and hand tracking, and timing, so hand-object contact and motion are preserved and ready to learn from, not reconstructed after the fact.
Rich language annotation — Chameleon already breaks each task into ordered steps, naming the tools and materials involved along the way. That structure becomes natural-language labels and step boundaries paired to the video, the grounding that language-conditioned and vision-language-action models need to connect instructions to behavior.
Ready for policy learning — The data arrives in formats aligned to how robot policies consume demonstrations. First-person frames, tracked poses, and language labels line up out of the box, so you can pretrain on human demonstrations and fine-tune across embodiments without rebuilding the pipeline first.
Request human capture from our creator network — Missing a skill, an environment, or an edge case? Commission it. Our global creator network captures the specific tasks and conditions your models lack, turning data collection from a months-long logistics problem into a request. Target the long tail, balance your distribution, or scale a whole new task family on demand.