AI Robot Brain PI Flexes General 2026 Home Assistant Tech (AI NEWS)

And in just a second, I’m going to show you an even more difficult long horizon task. But because of this long-term and short-term memory, the robot can carry out tasks exceeding 15 minutes in length. Let’s see if it burnt this bread here that it’s about to flip over having a little bit of trouble and it’s slightly burnt. But what is being solved here is the goldfish problem of robots where basically these robots cannot recollect what just happened. They don’t know the steps or how to tie them together and they get very confused over time.

So enter these two types of memory. Here you can see the robot is actually noting its task at the bottom. It says I’ve started the task. This is thanks to its short-term and long-term memory. First, let’s start with the short-term. These are stored observations from the processed video encoder and it interleaves the spatial and temporal attention of the robot to keep things efficient. Then it’s able to use its long-term memory to store these abstract textual notes in natural language that you see here. It says I’ve thrown the paper towel in the trash at the bottom of the sink and it will continue as it carries out this long horizon task.

But notice that it’ll start to actually compress its notes over time. And here’s the clever part. It’s able to not only compress its notes, but it’s also able to actively decide what to remember and how. And it uses its same reasoning process that it uses to pick its next sub-task. So it actually asks itself step-by-step, what should I do next and what’s worth writing down. This is part of this dual memory setup that can track the fine-grained recent actions as well as maintain a high level narrative of the complex multi-step task data at hand.

You see this robot cleaning up and compressing this entire paragraph of a report of what it’s doing. And this is where we’re about to enter in context adaptation. In the next clip, I’m going to show you, this robot can actually learn from mistakes in real time, which is one of the most impressive results of this MEM system. But while it’s washing this plate, it’d be interesting to know if this robot is waterproof. Can it get wet? And it would also be nice to see this work with dexterous fingers next time. But for now, just these two-fingered grippers and it’s washing this plate.

Trying these different strategies along the way to carry out this long horizon task of cleaning up this entire kitchen. But now let’s enter an unseen kitchen. So this is where this in context adaptation gets really important. This robot is learning in real time what to do and after failed attempts, it’s able to retry without any retraining. So this robot’s fumbling around for something in the refrigerator and it’s making these notes that it moved the potatoes to the countertop to maintain memory of where it is in this task to, I suppose, prepare a meal.

Tell me in the comments down below if you would trust this robot with preparing a meal for you or if it seems like it’s still a bit unsure. What’s really special about this is that the memory equips the robot to try in one direction and if it fails, it can try in another. And this is a really big deal because it means that you don’t need a model that never makes mistakes. You just need it to not make the same mistake twice. And that’s exactly where this in context adaptation really comes into play.

So this MEM system is actually a massive step forward for robots because now they don’t just follow single commands. They can manage an entire schedule of chores over hours, days, or even weeks. For instance, instead of telling the robot fold my laundry, you can instead say something like I get home at 6 PM, have dinner ready and clean the house on Wednesdays. And it would be able to follow this on a plan. And what’s really important here is that normally robots wouldn’t be able to plan and maintain memory across different timescales. But now with MEM they’re able to track all the way through and they’re even working on in the future, creating multiple timescales or different levels of abstraction, basically a hierarchy of memory from seconds to weeks.

But the biggest takeaway here from physical intelligence is that the bottleneck in robotics is quietly shifting from motor skills to cognitive architecture. So soon we’ll be able to get these home robots to just do whatever we want by simply speaking to them. And they’ll be able to reason and remember and be accountable to themselves. But tell me down below what other tasks or chores would you like to see this robot like this that maybe it takes a little while to complete tasks and end. But if they do get done specifically, how many dollars per hour would you be willing to pay for this robot to fix your food, clean up the house or comment down below if you think that there’s something even more important that this robot could be doing.

Pizza is definitely important. Anyways, like and subscribe and check out the rest of these videos here. Some of these other long horizon tasks that this robot carried out and tell me what you think below. [tr:trw].