Lightwheel drops 100,000 hours of first-person video
Two bets on the robot data shortage landed in the same week, one open and one paid.
Lightwheel has released 100,000 hours of egocentric video (first-person footage, the kind you get from a head-mounted camera) and put it on HuggingFace. The clips cover a wide range of human tasks, from cleaning and tidying to assembly and construction to retail work. Lightwheel CEO @bgxc laid out the reasoning on episode 102 of RoboPapers, hosted with @micoolcho and Jiafei Duan.
The pitch for egocentric data is simple. It is cheap to collect, it captures the breadth and diversity of real human tasks, and it is true to real physics. That last part matters because it is the thing simulation keeps having to approximate. Jiafei Duan said the episode also gets into how Lightwheel's focus on egocentric collection connects back to simulation.
Chris Paxton called it a huge contribution to the community and a big step for embodied AI, and went further than that on where the field sits now.
Figure's Index numbers keep climbing
On the other side of the same problem, Brett Adcock posted fresh numbers for Index, Figure's crowd-sourced video network that we covered last week. Figure hit 69,900 weekly active users last week, and users are uploading 35 minutes of data every second across the world.
That is up from the 30 minutes per second Adcock cited when Index came out of stealth. At launch he also listed 16 million video uploads, 264,000 downloads, and $15 million paid out to contributors, alongside a commitment to spend $1 billion on data and compute over the next 12 months.
What was actually shown
Data volume, and only data volume. Neither post shows a robot trained on any of this footage doing a task. Lightwheel's release is downloadable right now, so anyone can check what 100,000 hours actually looks like. Figure's numbers are company-reported and the network itself is paid and closed, which is the real split here. One camp is betting that an open pile of human video is the fastest path out of the data shortage. The other is paying people to build a private one.
Either way, the bottleneck argument in robotics has shifted. A couple of years ago the complaint was that there was almost no data at scale. Now the argument is about who owns it and whether first-person human video transfers to robot hardware at all. Nobody in these posts has demonstrated that transfer yet.
it truly seems as if the end of the robotics data shortage is in sight
Last week we hit 69,900 weekly active users