Galbot has debuted its ET1 humanoid robot, nicknamed "Galaxy Star Kid," featuring an advanced three-part "Astrobrain" AI architecture that enables fluid physical motion, real-time dance learning, and body balance. Engine AI’s T800 humanoid has been deployed in a factory setting, using a vision-language-action model alongside three-finger dexterous hands to execute real-world sorting, obstacle detection, and heavy balancing tasks. American robotics company 1X released new footage of its Neo humanoid robot demonstrating smoother, first-take movement capabilities while executing squatting and walking commands. Meta released evaluation results for its Muse Spark 1.2 multimodal model, demonstrating high-level task planning paired with low-level physical execution for robotics, competitive visual reasoning, and a new "Wild Artifact Bench" evaluation system.
Is this the next TeslaBot killer? Galbot just unveiled its brand new ET1 humanoid with several
new features. So today we're taking a closer slower look at the preview trailer and then
some insane real world demos. But first let's look at the exterior of this new ET1 model from
Galbot. It's nicknamed the Galaxy StarKid and this is its high definition head. Then zoomed in to
what looks like a power bar built into its chest. You can see here some anti-pinch fabric around its
shoulders. Then there seems to be some pinch points potentially in its joints. Comment down below if
you think that this would be safe for children. This is a bit smaller than your average person,
about four feet judging by some of the demos. Look at the way that the legs are able to
articulate specifically in its feet. And in just a moment I will show you this robot training in
real world scenarios and then showing off its agility specifically with the way that its feet
are able to articulate to do such fluid movements. Here's a full body shot of the entire robot but we
still don't know how many degrees of freedom it features. But what's even more impressive than the
ET1's physical exterior is actually its AI brain. This is called the AstroBrain and it has three
components to allow the ET1 to watch, learn, and do like never before. And in this demo I'm about to
show you, you'll see the entire learning process for a whole dance routine that's quite impressive.
But it works with these three components. The first being the cerebrum of the AstroBrain and this
handles understanding and high-level planning by unifying two previously separate approaches into
one model which allows this robot to anticipate physical consequences of an action before it's
decided its next move. Then component number two is the brain bridge which connects the cerebrum's
high-level reasoning to the cerebellum's low-level motor execution. And this leads to component number
three the cerebellum or the AstroBrain WBC which manages low-level millisecond fast physical
coordination balance and motor skills across dozens of joints allowing the robot to dance
with next generation fluidity. Comment down below if you think that this could be the smoothest
robot dance so far. And this cerebellum actually has 80.4 million parameters and it was trained on
about 100,000 hours of diverse human motion data which allows it to really control its whole body
in a way that we haven't seen yet before. And they're saying that the price tag of this robot
is going to be somewhere between fifty to a hundred thousand dollars. But because it's such a small
robot, tell me down in the comments whether or not you think it will be less than fifty thousand
dollars and if there should be some type of a leasing policy for this robot. Next, NGIN AI's T800
was just deployed in a real world factory and here's some point of view footage showing its
dexterous hands sorting tasks for the first time ever. And we can see that these are three-fingered
dexterous hands and it's using its vision language action model as well as whole body dual arm
coordination to be able to recognize these items and then sort them properly. And along with this
pick and place task you can see here the robot also features visual recognition system not just
for components that it's working with in its environment but it can also detect whether or
not another robot is within its vicinity so as to not bump into it or other objects nearby.
And along with some impressive full body motion control here they also showed a demonstration of
this robot picking up a payload with one arm and balancing. You can see that it walks here
quite steadily but the question really is how much is inside of this briefcase? Is it five pounds?
Fifty pounds? And of course these robots can collaborate together and right now they're
already working side by side with humans. But another Tesla robot challenger from American
company One X just showed off its new video of the NEO being prompted to squat down and walk.
And this is a first shot take and the robot is getting smoother in terms of its movements but
comment down below and tell us whether you'd go for Galbot's ET1 or One X's NEO.
At the same time the biggest labs in the world are currently investing in the frontier of embodied AI
as Meta just released new evaluation results and demos detailing the multimodal capabilities of
its MuseSpark 1.2 including a robotic setup that splits reasoning the same way that we just saw
one high-level model planning the task and a separate model handling the physical execution.
In fact Meta is using a specialized MuseSpark variant as the planning and orchestration layer
for physical tasks. The system splits manipulation into two levels where a high-level planner handles
scene understanding, distinguishes between similar objects and breaks a goal into subtasks
while a separate MuseSpark vision language action model carries out each step. And demonstrated
examples include sorting a hairbrush from a makeup brush, packing lunch items into a box
and navigating an unstructured environment to locate a rubber duck. But robotics is only one
part of what this update covers because on general multimodal benchmarks the picture is competitive
rather than dominant. MuseSpark 1.2 runs neck and neck with GPT 5.6 Sol on ZeroBench which tests
general visual reasoning and comfortably outperforms Claude Opus 5 and Gemini 3.7 Flash.
Then on simple VQA, a visual question answering test, MuseSpark 1.2 is only beaten by Gemini 3.7
Flash. But Meta says the model's real strength shows up once it's allowed to use tools rather
than reasoning from a single glance. Letting it inspect visual inputs more closely before answering
produces a larger performance jump than the improvement seen in MuseSpark 1.1 under the
same conditions. And on visual coding the model converts images or video directly into working
code for web pages and games, then re-examines its own output to self-correct. Then on the
third-party Design Arena leaderboard, MuseSpark 1.2 currently ranks first for converting video
into a website, places close second for converting a single image into a website, and sits in third
for image to front-end conversion, trailing Kimi K3 and Claude Fable 5 in that category.
And when it comes to audio-visual understanding, MuseSpark 1.2 pairs video comprehension and dense
captioning with agentic tools including web development, real-time search, and spatial
grounding aimed at turning raw video input directly into actionable output rather than
passive description. And Meta also previewed Wild Artifact Bench, an internal evaluation built to
score agents across any task type, deliverable format, or verification method, instead of relying
on fixed correct answers. Meta says frontier models still struggle with real-world agentic
work, like producing production-level games, which the new benchmark aims to capture. It
compares outputs from a tested agent against a baseline agent, with automated and human judges
choosing a preference to generate win rates, which is an approach that Meta says broadens task
coverage since nothing needs a predefined right answer. And Meta has already released a handful
of example tasks from the benchmark publicly, including analyzing a cardiac ultrasound sequence
and building a driving simulator trainer. Anyways, like and subscribe, and check out
this video here for more of the latest in AI and robotics news, and thanks for watching.