Nathan Kau

Co-founder and CEO @ Dream Robotics

[Nathan Kau @NathanKau]

https://t.co/lRfTWh0mdX

[Nathan Kau @NathanKau]

Robotics has a teleoperation problem.

Teleoperation used to be the only way to collect high quality demonstrations to post-train your robot. There was no way around the slow task completion speed (often 5x slower than a person!) and $100+/hr operation cost.

That's not the case anymore. We're proving, along with Sunday, Generalist, Reward, Chestnut, and more, that collecting data directly from people wearing sensorized backpacks, caps, and gloves produces data that is not only much cheaper, but better than teleoperation data once you put in the good old-fashioned engineering. Unlike egocentric video, this kind of system turns people’s hands into robot look-alikes with grippers, mitts, or gloves that match the kinematics, contact surfaces, and sensing of the robot hand so it can be used for post-training.

At @dreamrobotics_ , we take sensorized human data a step beyond. Our wearable is intuitive and dexterous, so much so that workers in real businesses can collect data on the job. Plus, patent-pending tactile, spatial sensors (lidar), and 4K HDR cameras give the robots information that no one else is collecting. We do this all in a one-pound backpack that costs less than $400 total, less than the cost of a single Insta360 camera, which we could only achieve by building our own edge computers, tactile sensors, and cameras.

Huge shout-out to my team for making this possible!

#physicalAI #AI #robotics #DreamRobotics #data #teleoperation #posttraining

[Nathan Kau @NathanKau]

@Novaris12138 Exactly!

[Nathan Kau @NathanKau]

@laneburgett @ActorLabs Awesome! Can you give a sense of the scale of the data you've collected?

[Nathan Kau @NathanKau]

Fantastic post! Dexterous versions of UMI are super interesting when they can be deployed into real facilities and worn by people doing their normal work. Data collected that way is naturally fast and done the way businesses want it done. Teleoperation and the policies produced are naturally slow, often 5x slower than a person!

[Nathan Kau @NathanKau]

Last November, Gabrael Levine, Alex Hattori, and I created Dream Robotics with the vision of building helpful robots for every business and every home.

From our work developing AI for @Tesla Optimus, building frontier fulfillment robots at Nimble, exos at @Google, and creating athletic quadrupeds for package delivery at LastFeet, we knew that to realize this vision we needed to solve data scaling and build robots that are actually cheap from day one. That’s why we created DreamOne, our inexpensive yet highly capable two-armed mobile robot, and DreamCatcher, a wearable for collecting high quality data from human work.

We’ve come a long way since our start in my parents’ garage in Palo Alto, house in Redwood City, and finally our bustling office in Mountain View. We now have an amazing team including Nicholas Robles, James Wang, Noah Baum, Eric Anderson, Anthony Andrews, and growing. Stay tuned for more!

#PhysicalAI #Startups #UMI #Wearable #Robotics #Stanford #MIT #Nimble #Google #Optimus #AI #DreamOne #DreamCatcher