Are Robotic Arms with Embodied Intelligence Easier to Implement Than Humanoid Robots?

In 2024, embodied intelligence and large models become hot topics. Many believe that the humanoid robot industry is on the brink of a “boom”, but how long will this night last? Another significant carrier of embodied intelligence—robotic arms—shows a dual strong prospect from commercialization to the application of new technologies in embodied multimodal large models.

Source｜Deep Blue Academy Group

—

While humanoid robots are gaining attention, robotic arms are the ones making money.

In 2024, as the curtain rises on technological transformation.

There are large models gaining momentum but engaging in price wars, and there are intelligent driving technologies that have almost fallen from grace and are now closed off to ponder new technologies.

And there is a rising star, predicted to be the “Ziweixing” of the year—humanoid robots.

Humanoid robots are like the protagonists of this drama, quickly capturing public imagination and media spotlight upon their appearance.

However, amidst this fervor, bubbles, doubts, concerns, and voices of pessimism also flood in.

Because when the spotlight dims, everyone turns to the capital market’s ledger.

“Oh no~”!

It suddenly becomes clear that the truly stable profit-makers are the seemingly ordinary—robotic arms.

Historically, profit-driven capital markets begin to take notice.

What have robotic arms been doing behind the scenes?

To say that a few years ago, the most “sensational on the internet” in the robotic arm field was the UP master Zhihui Jun (Peng Zhihui), who had not yet secured 1 billion yuan in cumulative financing.

At that time, he uploaded a video on Bilibili, showcasing the process of using his homemade robotic arm to sew grapes remotely in a bathroom! The barrage of comments was lively.

Three years later, the once “youth of Huawei” Zhihui Jun has now become the CEO of the hot humanoid robot company “Zhiyuan”, holding over 1 billion in financing.

The development prospects of robotic arms are increasingly promising.

For instance, at the recent World Robot Conference, there were demonstrations of using depth cameras combined with robotic arms for spatial recognition and object tracking, and others using facial tracking technology for automatic control of robotic arms…

Indeed, a strong driving force behind this is the rapid development of [embodied multimodal large models].

More and more work in the field of robotics is attempting to construct embodied multimodal large models to endow robots with high-level reasoning and low-level control capabilities from end to end.

Are Robotic Arms with Embodied Intelligence Easier to Implement Than Humanoid Robots?

So why have robotic arms been able to apply embodied intelligence before humanoid robots?

There are many reasons; the robotic arms have indeed not been idle over the past few decades, for example:

Clearer application scenarios: Compared to humanoid robots, robotic arms have clearer and broader application scenarios in industrial automation, logistics, and service industries, and they also have a price advantage.
Mature technology system: The technology of robotic arms has developed over decades, forming a relatively mature technology system and industrial chain, which is more conducive to integrating embodied multimodal large models.

More competitive cost pricing: The R&D and manufacturing costs of humanoid robots are much higher than those of robotic arms, especially in achieving complex human movements and perception capabilities. The high costs result in a very long investment return cycle and profit cycle for humanoid robots.
Greater policy support: From automobile manufacturing to electronic product assembly, robotic arms have become an indispensable part of industrial automation due to their high efficiency, precision, and reliability, with increasing policy support.

Therefore, compared to the current humanoid robot track, “robotic arms” have become a more stable and attractive business choice.

Moreover, the future, which is believed to be even more popular, “humanoid robots” cannot do without the support of “robotic arms”.

Thus, let’s take a good look at modern “robotic arms” and catch the wave of capital.

—

Intelligent Robotic Arms Emerge: How to Solve Technical Issues

In this series, we will core discuss several technical issues of the promising “intelligent” robotic arms:

One is the generalization issue of constructing the “base model” of robotic arms and the motion control solutions in real scenarios.
Another is the current issue of integrating the popular “fusion of vision, language, and action” embodied multimodal large models.
Lastly, we focus on exploring how robotic manipulation can approach human capabilities.

Of course, to delve deeper into these topics, a dedicated group has been established, currently with nearly 200 like-minded friends joining. If you’re interested, feel free to join us for discussions.

1. Embodied Multimodal Large Models

With the rapid development of embodied multimodal large models, big data, cloud computing, and other technologies, robotic arms are undergoing a profound transformation from single-function to multifunction, from fixed scenarios to flexible applications, moving towards smarter and more complex directions.

This presents new challenges and opportunities for algorithm technicians and the entire industry.

The integration of embodied multimodal large models with robotic arms, effectively pushes the boundaries of robotic arms in executing complex tasks.

Modern intelligent robotic arms are expected to land first in scenarios such as industrial manufacturing, flexible logistics, and commercial services, such as intelligent sorting and handling in automated factories, and executing natural language commands in household service robots.

The “application fields” of robotic arms are further expanded.

But! It is well-known that the current combination of multimodal large models and embodied intelligence has gradually risen from the most direct perceptual level to more complex issues of efficient planning, control, reasoning, etc.

How to more effectively integrate and micro-tune multimodal information for specific robotic tasks, and how to enhance the efficient reasoning and manipulation capabilities of embodied multimodal large models, has become a significant research direction.

In response to this issue, Peking University’s HMI Lab has developed the RoboMamba model, innovatively integrating a visual encoder with an efficient state space language model (Mamba), constructing a new multimodal large model that possesses reasoning capabilities for visual common sense tasks and robot-related tasks.

In September, the first author of this work, Dr. Liu Jiaming from Peking University (winner of the ICCV 2023 Continuous Generalization Learning Competition, ICCV 2023 Robust Multitask Learning Competition, and CVPR 2023 Autonomous Driving Occupancy Prediction Competition bronze medal).

Will specifically share on the issue of “embodied multimodal large models” in the efficient reasoning and manipulation of robots.

Friends interested in Dr. Liu’s research can join the group for discussions, after which Dr. Liu Jiaming will conduct an online public sharing to explore this cutting-edge issue together.

“Robot Grasping and Manipulation” Theme Exchange Month🦾

Are Robotic Arms with Embodied Intelligence Easier to Implement Than Humanoid Robots?

2. Generalizable Robot Manipulation

In the field of robotics technology, achieving generalizable robot manipulation has always been a hot research topic and challenge.

Generalizable robot manipulation refers to the ability of robots to autonomously understand and execute tasks based on existing knowledge and experience when facing unseen objects, environments, or tasks. The emergence of kinematics-guided prompt learning frameworks and embodied large models (like ManipLLM) enables robots to understand complex instructions, predict physical actions, and execute tasks in real environments.

However, to achieve generalizable robot manipulation, it often requires relying on massive data for imitation learning. However, collecting sufficient robotic data in real scenarios is prohibitively expensive, leading to significant cost issues.

For the “generalizable robot manipulation issue”, Dr. Xia Wenke from Renmin University of China’s Gao Ling Artificial Intelligence Institute will conduct a focused sharing. (Entry details at the end)

3. How Robot Manipulation Capabilities Approach Humans

Over the past few decades, robotics technology has made significant progress, and many commercial robotic platforms have been deployed, showcasing high precision and repeatability.

However, this exceptional precision often fails to effectively translate into the ability to manipulate many everyday objects. Currently, robotic object manipulation performance still cannot match that of humans.

Dr. Zhao Chao from the Hong Kong University of Science and Technology believes that breakthroughs in future object manipulation lie in achieving a kind of “imprecise dexterity”—even in the presence of imprecision in control, movement, and perception, robots can still achieve dexterous manipulation through robust and adaptive behavior.

This issue is also worthy of in-depth exploration and analysis by technical researchers. In this session, Dr. Zhao Chao (who studied under Professor Chen Qifeng) will specifically share on this topic.

(This group will operate for about 30 days, inviting 6 guests for in-depth sharing, currently open for free.)

—

Delving into Robotic Arms, Integrating New Technological Innovations

Thanks to the enormous potential market, broad application fields, mature technology systems, and strong policy support exhibited by embodied intelligence in emerging technology fields, the “robotic arm” track is heating up again, with its commercialization prospects even surpassing the much-anticipated humanoid robots.

Therefore, this month, the Deep Blue Academy has specially organized a “robotic arm” theme month. In addition to establishing a discussion group for everyone to explore, six researchers focusing on different subfields have been invited to conduct six live sharing sessions to promote domestic research and understanding of this theme. (Free access)

This content covers:Generalizable object manipulation strategies based on base models, interactive robotic arm motion control, achieving complex robotic manipulation with a single RGB camera, efficient reasoning and manipulation of embodied multimodal large models, and finally how robotic manipulation capabilities approach humans.

(Scan to join the group⬆️)

We believe that excellent people will eventually meet.

See you in our discussion group! Join us for live chats!

······ END ······

At that time, he uploaded a video on Bilibili, showcasing the process of using his homemade robotic arm to sew grapes remotely in a bathroom! The barrage of comments was lively.

Related posts

Leave a Comment Cancel reply