Disclosure: I work with a team that collects and licenses paid robotics training datasets.<\/strong><\/p>\n I’ve been speaking with robotics teams about human demonstration data<\/strong>, and every team seems to evaluate it differently.<\/p>\n Some only need egocentric video<\/strong>, while others require synchronized wrist views, task labels, collection metadata and licensing documentation<\/strong>.<\/p>\n We currently have small evaluation samples from the US, UK and Australia<\/strong>, covering:<\/p>\n \u2022 Egocentric demonstrations<\/strong> The small evaluation samples are free, but the complete datasets and custom collection services are paid<\/strong>.<\/p>\n For teams working on robot manipulation or embodied AI<\/strong>, what do you normally check first?<\/p>\n Camera coverage, task diversity, collection country, metadata quality or licensing?<\/strong><\/p>\n I’m mainly trying to understand what makes a sample genuinely useful before preparing more of them.<\/p>\n<\/div>\n submitted by \/u\/WideAmbition1964 <\/a>
\u2022 Egocentric + two wrist views<\/strong>
\u2022 Task and step labels<\/strong>
\u2022 Country and collection metadata<\/strong><\/p>\n
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
