Oh my God, the humanoid robot has evolved again

Mondo Culture Updated on 2024-02-27

The power of the machine reports.

Edit: sia

It looks like Moravec's paradox could be wrong?

It's only been two months since the Tesla Optimus evolved again.

In the latest**, the Optimus pace has reached 06 meters, while only half the pace of a healthy adult, has increased by more than 30% compared to last December**.

*Link: Optimus walks around the factory where it is produced every day, with a more steady pace and a more confident appearance** also in the bends). In addition, a slight torso and arm sway has been added. To do this, the team not only improved the human-like vestibular system, an important receptor for maintaining human balance and located in our inner ear, but also optimized the logic of the foot-end trajectory and contact with the ground, upgraded the motion planner and reduced the time required from input to output.

Here's a comparison of the Optimus walking movements from May and December last year to February this year**, where the evolution of speed and posture is visible to the naked eye.

*Link: Optimus has evolved, and it is still inevitable that netizens will spoof, is this pace quite Biden-style?

Not long ago, another high-profile humanoid robot company, 1x, also released its latest**. Over the past year, they've built a data engine to solve common mobile operations tasks in a fully end-to-end manner. The latest** showcases the new skills that EVE has mastered in this technology path.

This is a routine work in the logistics scenario, putting the packages on the shelves on the conveyor belt, selecting items and putting them in boxes, walking freely, and entering and exiting from the first room.

Skills for home scenes, tidying up ground items.

Eve also recharges itself, plugging and unplugging itself.

Each action is controlled by a 10Hz visual neural network that controls the robot's advance, arms, torso, and head. In order to prove that this nearly three-minute new ** is not boring (no remote control, no CGI special effects, no script replay), they used a one-shot shooting method.

A complete home and office environment is challenging because they are unstructured and constantly changing. Robots can't make too many assumptions about the world beyond the body, so 1x pursues the most versatile approach to autonomy.

They assembled a high-quality, diverse demo dataset across 30 EVEs and used it to train a "base model" that could understand a range of human activities and skills—from cleaning to tidying up the house, from picking up objects to interacting with humans and other robots.

Next, they fine-tune this base model to get a model that implements more specific functionality. For example, a model of general door operation and another model for warehouse tasks, and further fine-tune these models to solve specific tasks, such as opening this specific door.

Based on this strategy, EVE's new skills can be mastered in just a few minutes with just a few minutes of data collection and desktop GPU-based training. What's more, training eve doesn't need to be written, only data. This means that the average customer without an AI engineer can train new skills in humanoid robots.

The head of the 1x AI team also praised Tesla's latest progress.

According to Bernt Bornich, CEO of 1X, EVE will continue to evolve over the next four weeks, learning to tidy up the office and working on logistics on the production line of EVE and NEO humanoid robots. All of this uses the same data collection and training process as in the previous year.

Today, EVE has been deployed in some of the largest enterprises. Digit, a humanoid robot, is being tested at Amazon's R&D warehouse, as is GXO Logistics, the world's largest pure-play contract logistics provider, doing a similar job — working with autonomous mobile robots (AMRs) to transport containers of packages and items.

Figure AI, a humanoid robotics company in which Nvidia, OpenAI, and Bezos have invested, has announced a partnership with BMW to bring robots to automotive manufacturing, but has yet to disclose any concrete progress. In China, UBTECH's industrial version of the humanoid robot Walkers has entered the first workshop of NIO's second advanced manufacturing base, receiving on-site "training" and collaborating with humans to complete assembly and quality inspection operations on the automotive work line.

*Link: Industrial version of the humanoid robot Walkers, height 17 meters, replaceable end effector, equipped with 41 high-performance servo joints and multi-dimensional force sense, multi-eye stereo vision, omnidirectional hearing and inertia, ranging and other all-round perception systems, with a comprehensive upgrade of visual positioning and navigation, hand-eye coordination operation, gait control, multi-modal path planning vertical model and other technologies, the robot's autonomous movement and decision-making capabilities have been greatly improved, and accurate and safe synchronous operations can be achieved in the factory assembly line.

In addition to the existing players, just when the net worth skyrocketed and exceeded $2 trillion, Nvidia also announced the establishment of the generalist embodied agent research laboratory Gear (Generalist embodied agent research, which means "general embodied agent research") to enter the embodied AI.

We believe that in the future, every moving machine will be autonomous, and robots and simulations will be as ubiquitous as iPhones. Jim Fan, an AI scientist at NVIDIA and co-founder of the lab, said on X, "We're building a foundational agent: an AI with general-purpose capabilities that can learn how to act expertly in multiple worlds, both virtual and real." 2024 will be the year of robotics, the year of gaming AI, the year of simulation. ”

He also added an X saying: Our team is probably the richest embodied intelligence lab in the world.

Reference Links.

Related Pages