Surpassing Tesla! Momenta announces its entry into the era of autonomous driving 3.0

The development speed of Haomo Zhixing seems to be faster than we imagined...

September 13th is the 1020th day since the establishment of Haomo AI, and also the day of the sixth HAOMO AI DAY. On this special day, Haomo's autonomous driving development has shown new results.

With curiosity and confusion, let’s take a look at what technological advances in autonomous driving were reported at Haomo’s sixth AI Day.

Momenta’s Report Card from 0 to 1

The most intuitive and visible aspect of the industry’s perception of Momenta’s “speed” is the mass production of its autonomous driving system.

The HPilot intelligent driving system currently covers nearly 10 models and tens of thousands of vehicles under Great Wall Motors, including popular brands such as Tank, Wei, Haval, etc.

Therefore, the first "first" of Momenta is the scale of mass production coverage. Whether it is Tesla, new forces, or other autonomous driving companies, achieving tens of thousands of installations in about three years is unprecedented.

In addition, in terms of terminal logistics automatic delivery, Haomo released the second-generation terminal logistics automatic delivery vehicle - Xiaomotuo 2.0 in April. It is the industry's first 100,000 yuan-level terminal logistics automatic delivery vehicle for the commercial market. Currently, Xiaomotuo 2.0 is being delivered to customers one after another.

In terms of AI artificial intelligence technology, based on the most advanced international AI technology concepts, Momenta launched the country's first intelligent driving data intelligence system MANA in December 2021. Currently, the MANA data intelligence system has become the core driving force for all Momenta product iterations.

Zhang Kai said that in the past 1,000 days, Haomo Zhixing has just crossed the life and death line of a startup from 0 to 1. Now, Haomo is entering a rapid development stage from 1 to N.

MANA's Six Major Milestone Upgrades

Urban navigation assisted driving scenarios are the core breakthrough point of current autonomous driving functions and are also a battleground for strategists.

However, the technical difficulty faced by autonomous driving systems can be said to increase exponentially from high-speed scenarios with simple road and traffic conditions to urban scenarios with a large number of traffic participants and extremely complex road and traffic conditions.

At present, there are "four types of scenario problems and six major technical challenges" in urban roads. The scenario problems mainly include "urban road maintenance", "dense large vehicles", "narrow lane change space" and "diverse urban environment".

To solve the above scenario problems, the technical level faces six major challenges: how to more efficiently convert data scale into model effects, how to make data more valuable, how to use re-sensing technology to solve the problem of understanding real space, how to use the interactive interface of the human world, how to make simulation more realistic, and how to make the autonomous driving system move more like a human.

In order to meet the above challenges, MANA's perception intelligence and cognitive intelligence have been updated and upgraded:

1. Enable self-supervised learning

MANA creates model effects by using a self-supervised learning method with unlabeled data from large-scale mass-produced vehicles. Compared with training with only a small number of labeled samples, the training effect is improved by more than 3 times. This allows the data advantage of Haomo to be efficiently converted into model effects to better adapt to various perception task requirements of autonomous driving.

2. Construct an incremental learning and training platform

During the training process, Haomo does not use all the existing data because it is expensive and slow. Instead, it extracts part of the existing data and adds new data to form a mixed data set. During training, the output of the new model and the old model should be kept as consistent as possible, and the fit to the new data should be as good as possible.

Compared with conventional methods, Haomo can save more than 80% of computing power to achieve the same accuracy, and the convergence time can be increased by more than 6 times.

3. Use temporal transformers to provide real-time spatial cognition capabilities

By using a time-series transformer model to perform virtual real-time mapping in the BEV space, the output of lane line perception is made more accurate and stable, allowing urban navigation and autonomous driving to say goodbye to dependence on high-precision maps.

4. Use the human world’s interactive interface to perceive the world more accurately

At the last AI Day, Haomo introduced how to solve the interaction problem between the autonomous driving system and traffic lights in urban environments without relying on high-precision maps. Recently, Haomo is upgrading the perception system on the car, hoping to add special recognition of the vehicle's signal light status, including brake lights and turn signals.

In this way, Haomo can drive more safely and comfortably in scenarios such as the vehicle in front slows down or surrounding vehicles cut in.

5. Use traffic flow simulation

Faced with intersections, the most complex scenes in the city, MANA introduced high-value real traffic flow scenes into the simulation system, and cooperated with Zhejiang Deqing and Alibaba Cloud to introduce intersections, the most complex scenes in the city, into the simulation engine, build an autonomous driving scene library, and verify the real simulation of autonomous driving with higher timeliness and more realistic micro-traffic flow, effectively solving the "long-standing" problem of passing through urban intersections.

6. Learn common sense and anthropomorphize actions

By deeply understanding the massive amount of human driving across the country, learning common sense and anthropomorphizing actions, Haomo's assisted driving decisions are more like actual human driving behavior. It can choose the optimal route based on actual conditions to ensure safety, and the feeling is more like that of an experienced driver.

Laying out cutting-edge technology

Go all out to sprint towards the era of autonomous driving 3.0

What is the era of autonomous driving 3.0?

Gu Weihao believes that the development of autonomous driving technology in the past decade can be divided into three stages: the earliest hardware-driven method, which is the autonomous driving 1.0 era; the software-driven autonomous driving 2.0 era; and the data-driven autonomous driving 3.0 era.

As a new trend in the current development of AI, the Attention Big Model has brought opportunities and challenges, becoming one of the important driving factors in the era of autonomous driving 3.0.

The most significant feature of Attention is its simple structure. Basic units can be infinitely stacked to obtain a model with a huge number of parameters. With the increase in the number of parameters and the improvement of training methods, the effect of large models has surpassed the average human level in many NLP tasks.

However, the large Attention model also faces a major challenge. Its demand for computing power far exceeds Moore's Law, which makes the training cost of the large model very high and makes it very difficult to implement it on terminal devices.

However, during Momenta's research and development process, the large transfomer model based on the Attention mechanism has challenges such as high computing power requirements, high training costs, and high implementation difficulty, as its demand for computing power far exceeds Moore's Law.

In this regard, Gu Weihao, CEO of Haomo Intelligent Driving, said: "Haomo is reducing the cost of autonomous driving through low-carbon supercomputing, implementing large models on the vehicle side by improving the design of vehicle-side models and chips, and making large models more effective through data organization."

Currently, Tesla is leading the world in entering the era of autonomous driving 3.0. At this year's Momenta AI DAY, Gu Weihao said that Momenta is most likely to become the first Chinese company to enter the era of autonomous driving 3.0, and is making every effort to achieve this goal.

Summarize

While most players in the market are still focusing on the 2.0 era, Momenta is ready to sprint into the 3.0 era.

In the arena of autonomous driving development where experts gather, courage alone is far from enough. The choices made by Momenta are all based on its own strength, and the "Momenta Model" may also become a new paradigm for the development of autonomous driving in China.