Robot + LLM ≠ Embodied Intelligence?

Heart of Machine PRO · Member Newsletter Week 36

---- This week we will explain to you ⑤ important things in the AI ​​& Robotics industry that are worthy of careful review ----

1. Robot LLM ≠ Embodied Intelligence?

What is the next step in the LLM technology roadmap for universal humanoid robots? What are the major technical challenges in general robot LLM leading to embodied intelligence? Before LLM became popular, how did Boston Dynamics make robots? What opportunities will breakthroughs in scene understanding and human-machine collaboration technologies bring? ...

2. Is the open source ecosystem of Llama 2 a pie or a trap?

Is the open source ecosystem brought by Llama 2 reliable? Is Baichuan-2 expected to be the domestic replacement for Llama 2? What is the significance of open source LLM training slices? Open source and closed source, what is the competitive landscape in the domestic large model field like? Is Llama 2’s open source ecosystem reliable? Could Baichuan-2 be a domestic alternative to Llama 2? What is the significance of open source LLM training slices? What is the competitive landscape between open source and closed source in the domestic large model field?

3. Is RLAIF a reliable alternative? Replacing humans (H) with artificial intelligence (AI)?

How is RLAIF implemented? How does AI annotation enhance RL? What are the advantages of RALIF? How does LLM performed based on RLAIF training? Is it feasible for RLAIF to replace RLHF? Will RLHF be needed in the future? What other recent RL research is Google doing? ...

4. OpenAI’s secret training GPT-5

GPT-5 Is there any gossip? What is the function of GPT-5? Does GPT-5 really exist? Sam Altman said before that he didn’t work on GPT-5? ...

5. How many years did it take for AI to take over the translation work?

Why were all the Spanish website editors “layoffs”? Is it reliable to use AI to translate websites? Learn about the history of AI translation development starting from Google? Do you remember what AI translation looked like ten years ago? Where will AI translation develop now? Why are all the editors of Spanish-language websites fired? Is using artificial intelligence to translate websites reliable? Let’s take a look at the development of artificial intelligence translation starting from Google. Do you remember what artificial intelligence translation looked like ten years ago? In which direction will the current artificial intelligence translation develop?

This full version of the newsletter includes 5 topic interpretations and 29 important news updates on the AI ​​and robotics track. Among them, there are 9 technical points, 11 domestic points, and 9 foreign points

This newsletter has a total of 24646 words, and you can try it for free up to 7%

You only need to consume 99 WeChat beans to redeem the complete interpretation of this issue, which is equivalent to RMB 9.9

Interpretation of important things ①Robot LLM ≠ Embodied intelligence?

Time: September 6

Event: Zhihui Jun recently revealed in an interview that his entrepreneurial team’s general humanoid robot LLM development plan includes the establishment of a data center and iterative reconstruction of the hardware structure.

Zhihui Jun, what are your thoughts on the next step in general humanoid robots and LLM technology?

1. Zhihui Jun said in the interview that in the embodied intelligence technology route of LLM universal humanoid robot, the core threshold lies in data. One of the recent focuses of Zhiyuan Robotics is to build its own data center.

Zhihui Jun summarized that his data work will involve "supervised learning data", "simulation data" and "AIGC generated data"

Zhihui Jun said that the next plan is to launch Lingang within a few months and establish a scenario and simulation platform to fill in motion data to enhance the generalization ability of the robot

2. Another focus of Zhiyuan Robot’s work is to iteratively reconstruct the hardware structure with the goal of enhancing the robot’s motion performance.

Zhiyuan Robot currently states that the price of humanoid robots will be controlled below 200,000 yuan

Mr. Zhihui said that if the price cannot reach 200,000 yuan, the humanoid robot will not be commercialized

② The valuation of 200,000 yuan can be compared with the 1-2 year investment return period required for robots to replace some workers in the new energy automobile manufacturing industry.

4. The Zhiyuan Robotics team’s method of controlling costs for mass production involves two aspects:

Adopting the self-developed route, such as self-developed core components such as joint motors and dexterous hands, can reduce costs by half

Reduce hardware costs by using software and algorithms to meet accuracy requirements

Zhihui Jun said that their primary goal is to achieve commercialization in the field of industrial manufacturing, and they plan to achieve this goal in the second half of next year

6. Zhihui Jun also mentioned a hidden line in the company's commercialization, namely: "laying eggs along the way" on the way to the ultimate goal of universal humanoid robots.

① Universal humanoid robots involve the most comprehensive robotic technology stack, and their implementation process involves the development and optimization of a variety of cutting-edge technologies, which can give rise to a variety of innovative robot products in specialized forms.

In addition to Zhiyuan Robot’s Expedition A1, what other teams are developing universal humanoid robots in China? [6] [7]

Are universal robots and LLM equivalent to embodied intelligence? [2] [3] [26]

Yao Qizhi, winner of the Turing Award, academician of the Chinese Academy of Sciences, and director of the Institute of Cross-Information at Tsinghua University, said at the 2023 World Robot Conference: Future AGI needs to have embodied entities that interact with the real physical world to complete various tasks. Such tasks can bring real greater value to the industry. At the same time, Yao Qizhi pointed out that embodied robots currently encounter four main challenges:

1. Robots cannot have a basic large model like a large language model to directly achieve the lowest level control in one step.

2. The challenge of computing power. Even with the Robotics Transformer model developed by Google, many improvements are still needed to achieve robot control

3. How to integrate all the multi-modal sensory perception of robots still faces many problems that need to be solved.

The development of robots requires a large amount of data collection, and also faces many security and privacy issues

How did Boston Dynamics make robots before LLM became popular?

In 2021, Pat Marion, a senior robotics engineer at Boston Dynamics and head of Atlas perception software development, published an article explaining the technology behind Atlas Parkour. [4]

Atlas's realization of excellent parkour capabilities mainly involves three aspects of technology: parkour cognitive capabilities, Atlas behavior library and model predictive control

2. Parkour cognitive ability: including the use of advanced depth cameras, perception algorithms and advanced maps and other components

① Atlas uses a TOF depth camera to generate a point cloud of the environment at 15 frames per second. The point cloud is a large-scale collection of ranging.

② TOF (Time of flight) literally translates as "time of flight". The ranging principle is to continuously send light pulses to the target, and then use the sensor to receive the light returned from the object, and obtain the target distance by detecting the flight (round-trip) time of the light pulse.

The above is the detailed content of Robot + LLM ≠ Embodied Intelligence?. For more information, please follow other related articles on the PHP Chinese website!