Elon Musk unveils the mind-blowing "Optimus Academy" where thousands of robots learn through real-world self-play and hyper-realistic simulations—could this close the gap to mass production? A powerful new Chinese humanoid stuns with 43 degrees of freedom, insane torque, stair-climbing agility, and dexterous hand skills that even humans struggle with—what happens when it links to your brain? Another humanoid just conquered a full parkour course using only onboard vision and real-time AI decisions, chaining explosive moves like a pro athlete—imagine these in real danger zones. Google drops Gemini 3.1 Pro, more than doubling reasoning power on tough new tests and powering everything from creative projects to agentic tools—is this the core upgrade AI needed?
Is 2026 the year of Tesla's Optimus, or something else? Today on AI News, Elon Musk reveals the
Optimus Academy for robots, plus we take a closer look at several new humanoids to compete with it.
So first, here's Elon setting up the training grounds that Tesla's using.
So you have a few tens of thousands of humanoid robots doing different tasks,
and then you've got, you can do millions of simulated robots in the simulated world,
and you use the tens of thousands of robots in the real world to close the simulation to reality gap.
Then, he talks about an Optimus Academy of robots that learn by playing with themselves.
For the robot, what we're going to need to do is build a lot of robots and put them in kind
of like an Optimus Academy, so they can do self-play in reality. So we're actually pulling
that out, so we're going to have at least 10,000 Optimus robots, maybe 20 or 30,000 that can do,
that are doing self-play and testing different tasks. And then Tesla has quite a good
reality generator, like a physics-accurate reality generator that we made for the cars,
we'll do the same thing for the robots, and actually have done that for the robots.
So how reliable is Elon's production timeline, given Tesla's current tech?
A bunch of the things that we've done for the car are applicable to the robot,
so we'll use the same Tesla AI chips in the robot as the car, we'll use the same basic principles,
it's very much the same AI.
And you can check out the rest on Dwarkesh Patel's channel, but competition is heating up,
as Xhumanoid just released this demo of its Tiankong 3.0 robot,
now with 43 degrees of freedom throughout its body.
And in just a moment here, I'll show you the dexterous demo with its hands,
but peak torque of 400 newton meters, which is on the upper end of robot power,
with a relatively small frame, 169 centimeters, and the running gate looks pretty realistic.
And we can see here that it's able to coordinate its entire body,
using its embodied AI, and in just a moment we'll see how it works with its hands.
Look at the way that it can step up those stairs, that's pretty impressive.
And jump right back down, and here's for the dexterous demo,
this is something that most humans can't even do.
Would be good to see this in one straight shot with no editing.
And it catches the ball there, and then it gets into a real world work demo.
So let's see what this robot can do in the real world, beyond just theatrics.
Here we see it turning knobs and pressing buttons.
And you can also use its vision language action model to pick up items, clear spaces,
as well as handle high payload. Look at the way that it's able to pull this cart
while maintaining its balance. And what's interesting here is that they say that they're
going to provide a demo soon with some kind of brain computer interface, or at least that's what
looks like from this preview. So would you hook up Tenkong 3.0 to your brain?
Meanwhile, a humanoid robot just completed a parkour course using only its onboard depth
camera and a simple velocity command. This is thanks to perceptive human parkour, or PHP,
which is a brand new framework that enables the Unitrix G1 humanoid to autonomously chain
together dynamic movements, stepping over obstacles, climbing onto platforms, vaulting
barriers, and rolling off of ledges all in real time with no pre-planned paths.
And the challenge here goes way beyond stable walking, because parkour demands explosive,
highly dynamic movements that push the limits of balance and coordination.
And it also requires that the robot is able to perceive its environment, decide which skill to
use, and execute that skill while preserving the fluidity of human motion. And while most
systems can do one of these things, doing all of them together autonomously is the hard part.
And this is where PHP comes in, solving this with a modular approach built on motion matching.
So the system stores a library of retargeted human parkour movements as atomic skills,
then uses nearest neighbor search in a feature space to compose these skills
into long horizon kinematic trajectories. And this allows the robot to smoothly transition
between movements without the jerky handoffs typical of state machine approaches. So once
these kinematic trajectories are composed, reinforcement learning expert policies are
trained to track each motion. And in the end, it results in one unified policy that handles
multiple skills, takes depth images as input, and makes context-aware decisions about which
movement to execute based on obstacle geometry. And the results in the real world were striking,
as the G1 was successfully able to climb up obstacles that were 96% of its own height,
and it completed multi-obstacle courses with closed-loop adaptation, meaning that it adjusted
its behavior in real time when researchers moved or perturbed obstacles mid-run. And all of this
used onboard sensors with a discrete 2D velocity command, with no external motion capture,
no human in the loop, just perception, decision making, and execution as one continuous system.
Plus, Unitree also showed off this demo of its cluster cooperative rapid scheduling system,
which allows all of these robots to act as one coordinated system.
So what else could this be used for beyond entertainment that would be useful? Maybe
construction? Or cleaning up a dangerous site? Comment down below. And finally, Google just
released Gemini 3.1 Pro, and it's way better than Gemini 3.0 Pro in multiple areas. First,
on ArcAGI-2, a test that evaluates a model's ability to solve entirely new logic patterns
that it's never seen before, 3.1 Pro scored a 77.1%, which is more than double the reasoning
performance of Gemini 3.0 Pro. In fact, this follows a recent update to Gemini 3 DeepThink,
which targeted scientific research and engineering challenges, but Gemini 3.1 Pro is the upgraded
core intelligence underlying those capabilities, now shipping across Google's consumer and developer
products. And already, Gemini 3.1 Pro is available in the Gemini API and Google AI Studio, Gemini CLI,
Google's agentic development platform, Antigravity, Android Studio, Vertex AI, Gemini Enterprise,
the Gemini app, and Notebook LM. And on top of this, higher usage limits are rolling out for
Google AI Pro and Ultra subscribers. So all together, Google's 3.1 Pro is being positioned
for tasks where simple answers aren't enough, such as with synthesizing data into unified views,
generating visual explanations of complex topics, or supporting creative projects that require
multi-step reasoning. And while it's currently in preview, Google says they're validating updates
and pushing further on agentic workflows before general availability. Anyways, like and subscribe,
and check out this video here for more of the latest in AI and robotics news.