This post discusses a couple of important talks and keynotes from the Humanoids Summit underway this week. Please excuse typos, as I am trying to type this fast. There will be more content from the summit later this week as I find time to type these out.
Physical Intelligence
The summit started off from a keynote by Sergey Levine from Physical Intelligence. We have discussed the pioneering work from Physical Intelligence many times in this newsletter. Sergey’s talk was mainly on the 𝝅0 robot foundation model being developed by them. Sergey is a great speaker and covers a lot of territory in a 30 minute talk. Some things that stood out:
The recipe for success is a combination of pre-training and post-training steps. Pre-training is done on a very diverse dataset comprising of 10000 hours of data from a variety of robots… much of this data coming from the Internet or open source repositories. The pre-training phase lasted about 2 weeks. On the other hand, post-training data is high quality, narrow and specific to the robot for which the policy was being built. It is usually around 20 hours of data collected from the robot. By itself, training on per-training data or post-training data will not create a great model. It is the combination of pre training and post training data that is necessary and critical.
The robot acting on the model is far from perfect, makes a long list of mistakes, but per-training and post-training combination allows it to recover. Ability to recover from mistakes was unexpected, and what led to the robot making steady progress in what it could learn.
They demoed a cool video of what happens when a person messes around with the robot such as throwing a towel at it while it is folding clothes. The robot handles perturbations beautifully, and throws item messing with it in a bin.
Their next goal is to handle language interactions so that the robot can respond to a human talking to the robot. Once that happens, the robot can now get feedback from humans and correct itself.
Their goal is to get 𝝅0 working on more robots. They successfully deployed it on the Astribot robot. Developers of the model 𝝅0 never touched the Astribot robot, nor had access to any of its data. But the 𝝅0 model could be successfully trained for the Astribot, which made Astribot successful in preparing coffee.
Their next goal is to improve the phase of fine tuning the model, because fine tuning is still a very manual work. There is a lot of human legwork involved in fine tuning and adopting the model to another robot.
1x technologies
The next talk was from the CEO of 1x technologies, Bernt Bornich. 1x technologies is a startup focused on creating safe and intelligent humanoid robots. The company is co-located in Norway and the Silicon Valley. Some interesting things that came up during this talk:
Bernt believes that this is the right time in our lives where all the technologies that enable humanoids come together… Sensors, actuators, Artificial Intelligence, and software.
He sees a growing shortage of labor due to a shrinking workforce, particularly in critical areas just as helping the elderly or in nursing homes. He believes that safe and intelligent humanoids will go a long way in helping out in these areas.
While there are a number of companies developing humanoids, what matters most is who is going to achieve intelligence first… which depends a lot on a lot of data gathering capabilities. Diversity of data is very important. The data can then be leveraged for advanced models.
Nature is the biggest teacher and mimicking dynamics of nature is important. Hence 1x’s Neo home humanoid is designed to live around us. It will be passively safe, very quiet, and can blend in our homes. The secret to robotics is not to be stiff. Therefore, the Neo robot uses tendons, where as, other robots have gears and motors. There is a great video on this subject.
1x also had its robots for demo in the Exhibits area. While the Neo robot was not powered ON, two Eve Humanoid robots (which work on wheels instead of legs) were seen roaming around the Exhibit area. Here are some pictures of Neo and Eve.
I will post more tomorrow, with more material for paid subscribers. If you can afford one, please consider joining us. Many employers reimburse the cost of a newsletter subscription. Please contact me for any details.