Humyn Labs 获 4700 万美元 A 轮融资,用工厂实景数据训练机器人
机器人模仿学习最缺真实工厂场景数据,Humyn Labs 让工人戴采集眼镜在产线上录第一视角,再拆成 reach、grasp 等原子动作供策略训练,还自带 benchmark 验证。刚拿 4700 万美元 A 轮,做具身智能数据的朋友可以关注这条数据供应链思路。
机器人数据采集公司 Humyn Labs 完成 4700 万美元 A 轮融资。其核心做法是让工人在真实工厂车间佩戴自研的采集眼镜,录制第一视角操作视频,弥补机器人模仿学习数据大多来自整洁实验室、在昏暗杂乱环境中表现不佳的问题。
在数据处理环节,Humyn Labs 会先过滤掉看似正常但实际不可用的自我视角素材,再把留下的内容拆解为 reach、grasp、lift、place 等原子动作单元,称为策略可学习的最小单元。这种结构让同一个技能能跨场景泛化,使数据真正达到可训练状态而非仅是录制完成。
公司还配备自有研究团队在数据之上构建 benchmark,数据经测试后才交付。这种"采集—过滤—结构化—验证"的闭环被视为 Physical AI 数据规模化的可行路径。
Robots learn by copying people, but most of that footage comes from tidy labs, so they stumble on cluttered, dim factory floors.
@humynlabs records real workers on real floors, using capture glasses they built themselves. It’s interesting to see something like this in the data quality space.
A lot of egocentric data footage looks fine and still can’t be used. Humyn Labs filters that out before anything gets labelled.
What survives is broken into atomic primitives like reach, grasp, lift and place, which they call the smallest unit a policy learns.
My read is that this is what makes the data training-ready rather than just recorded, because a policy can learn the skill across different scenes.
They have their own research team building benchmarks on top of it, so the data gets tested and doesn’t just get shipped. That whole loop is a great solution to scaling Physical AI.
Congrats on the $47M Series A funding, looking forward to what’s next.
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Humyn Labs, Physical AI, robot learning, embodied AI.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).