10 hours ago
ARTICLE AD BOX
The Vice President of AI at 1X Technologies has shared a number of important updates connected their humanoid robot NEO.
While nan title to create humanoid robot is heating up, 1X is firmly focused connected delivering robots into nan home, addressing mundane tasks, while others attraction connected business applications.
In nan posts today, we spot NEO performing tasks for illustration picking up leaves and placing them into a bag. For those pinch trees that routinely suffer their leaves arsenic nan seasons change, this task is thing they’ll beryllium very acquainted with. Showing a robot performing tasks for illustration this showcases nan very human-like capabilities of their design.
When we spot videos of humanoid robots online, it’s important to understand precisely what we’re seeing. There person been videos wherever a distant usability is controlling nan movements of nan robot, nevertheless that’s not nan lawsuit here.
As Jang highlights successful his post, nan robot is operating 100% autonomously. I would person loved to spot nan robot locomotion complete to nan leaves, identifying nan heap of leaves, opening nan container and really it resolves nan container getting afloat aliases exceeding nan payload measurement nan robot is tin of carrying. Humans would quiet nan leaves into a bin, truthful while nan end-to-end task whitethorn not beryllium solved today, this is simply a promising motion for nan future.
In different post, we spot nan robot completing 1 of nan astir hated tasks successful nan home, stacking nan dishwasher. When we analyse nan complexity of this task, it’s easy 1 of nan much difficult to group a robot to.
Objects are different sizes and shapes and tin only fresh successful a group set of locations successful nan dishwasher drawers, it besides has to woody pinch objects that whitethorn beryllium slippery. The circumstantial demo is comparatively basic, successful that nan dishwasher is already unfastened and waiting for a cup to beryllium loaded.
For this to beryllium applicable successful each home, nan solution needs to beryllium generalised, accommodating for each nan different configurations of dishwashers, aliases spell done a learning shape astatine your home, pinch your circumstantial devices.
What 1X is attempting to show present is nan manipulation of objects, retrieving nan cup from nan sink, moving it betwixt hands to trim nan assemblage activity required. It besides looks down to nan tie to let it to spot nan cup successfully.
Again, it is important to see what’s not shown successful nan videos to admit nan advancement coming and nan remaining to-do list.
Finally we spot a 3rd video wherever NEO walks confidently complete to nan sofa and places a sofa cushion. The task whitethorn beryllium a small little common, though those pinch kids and pets, having cushions backmost successful their spot astatine nan extremity of nan time would beryllium a awesome help.
This task requires nan robot to understand what a cushion is, wherever nan sofa is and that nan cushion typically goes successful nan area of a couch. The robot leans complete to spot nan cushion down without toppling over, making nan basal adjustments to counter-balance nan weight.
These are promising signs that nan robot could 1 time cleanable your location pinch nan correct training. There’s besides rugs and different obstacles to woody pinch and erstwhile successful a shared environment, it’s important nan robot is resilient and tin beryllium move successful achieving it’s target. Imagine a pet aliases kid moves into nan target position for nan cushion, nan robot has to respond accordingly, for illustration placing it connected an alternate sofa aliases chair.
Yang shares a thread connected X that helps america amended understand what we’ve seen today.
Everything is learned from information – nan web besides controls afloat assemblage motion, including directing nan RL little assemblage controller + arms and spine joints astatine nan aforesaid time.
At @1x_tech we’re making humanoids for nan home. A short thread connected why we’re going for nan location marketplace first, arsenic opposed to last.
The location is nan “Final Boss” of robotics, because it is truthful radically unstructured. Every location is different and location are thousands of sub-tasks progressive successful thing arsenic elemental arsenic doing nan laundry (unscrewing detergent caps, unbuttoning shirts).
And yet, nan location is wherever you find nan eventual wealth of divers information basal to create a wide intelligence. Diverse information crossed galore tasks and environments is simply a basal constituent of making wide intent robots.
Most robotics companies presume that autonomy has to beryllium solved 1 task astatine a time, because autonomy is hard. Solve 1 task, and past usage nan money generated by that task to grow into much commercially valuable tasks.
However, adding capabilities this measurement is excessively slow – moreover if you onboard a caller task each quarter, you will only person 20 skills successful 5 years. Home chores request thousands of capabilities, not a handful.
In autonomous cars, immoderate thought that autonomous trucking connected nan road would beryllium overmuch easier than municipality driving because nan problem did not require wide intelligence.
Paradoxically, making a constrictive AI activity good successful moderately unstructured situations is measurement harder than making a wide AI OK astatine very unstructured situations. Moderately unstructured environments are much wide than you think.
You request to cod divers information successful bid for thing to work, moreover connected constrictive problems. However, it’s difficult to cod failures successful nan chaotic for thing arsenic high-stakes arsenic moving a truck. And nan “structured environment” really useful against you – uncommon events are moreover rarer truthful it’s difficult to excavation nan information that helps to make it safe.
That’s why Tesla’s AutoPilot ADAS strategy – trained for user – useful great, while location is still nary viable autonomous trucking solution today.
The aforesaid is existent for humanoid robots. One mightiness deliberation that solving mill activity is easier than going for autonomy successful nan home, but nan tasks progressive will still require a really generalized beingness knowing because they are adjacent to people, who are fundamentally unstructured.
ChatGPT showed nan world that autonomy does not person to beryllium solved 1 task astatine a time. It is an unthinkable business lawsuit study successful nan consumer-first distribution strategy for AI. Solve a wide group of tasks for user users, and past endeavor take will travel rapidly.
English (US) · 
Indonesian (ID) ·