GM’s Cruise, an autonomous driving subsidiary, has provided a detailed level of technology and deployment roadmap. This is intended to show how Cruise was built. Self-driving car It is safer and more scalable than any human-driven vehicle, including vehicles equipped with advanced driver assistance systems.
While Cruise clearly advocated its own technology (not to mention trying to hire fresh talent), the event was also a general self-driving car discussion. Each engineer or product lead who spoke on Thursday introduced a variety of components, from how to use simulations, developing their own chips and other hardware, to designing apps and the vehicle itself.
Comment The brand’s “Under the Hood” event built from the CEO Dan Aman Held last month on GM’s Investor Day, where he set out the company’s plans to launch a commercial robotaxi and delivery service. Start with a modified Chevrolet Bolt Eventually, it will expand to tens of thousands of dedicated Origin AV troops over the next few years.
cruise Just approved in California One permission away from being able to perform commercial delivery services and still be able to charge Self-driving ride hailing.. Still, Cruise believes it can reduce costs enough to scale up and out quickly.
Method is as follows.
Scaling using simulation as well as validating the system
Cruise relies not only to prove safety cases, but also to simulate to expand into new cities without first running millions of miles of testing.
The company needs to map the cities it enters. However, there is no need to remap cities to track inevitable environmental changes such as lane changes or street closures. When Cruise goes to a new city, it starts with a technology called WorldGen. This means generating the entire city accurately and on a large scale, “from quirky layouts to details,” so engineers can test new operational designs. Domains, according to Sid Gandhi, Technology Strategy Leader for Cruise Simulation. In other words, WorldGen is at the stage where future simulations are set up.
To ensure the creation of an optimal world, Cruise systematically measures light from various streetlights in San Francisco, taking into account lighting and weather conditions in 24 different and unique time zones.
“The combination of a loyal environment and procedurally generated cities unleashes the ability to efficiently expand our business into new cities,” Gandhi said.
Next, he laid out the “Road to Sim” technology. This transforms the actual events collected by the AV on the road into an editable simulation scenario. This ensures that the AV does not recede by testing against scenarios that the AV has already seen.
“Road to Sim combines perceptual information with millions of miles of real-world learned heuristics to recreate a complete simulation environment from road data,” Gandhi said. “Once the simulation is complete, we can actually create a permutation of events and change attributes such as vehicle and pedestrian types. This is very easy and very easy to build a test suite to accelerate AV development. It’s a powerful method. “
Morpheus is a specific scenario that Cruise could not collect under real road conditions. Morpheus is a system that can generate simulations based on specific locations on a map. Use machine learning to automatically populate the number of parameters needed to generate thousands of interesting and rare scenarios for testing AV.
“We’re working on a long-tail solution, so we’re less dependent on actual testing. If an event occurs that rarely happens, it can take thousands of miles to properly test and it’s not scalable. That’s it. ” Gandhi. “Therefore, we are developing a technology to scalablely explore a large parameter space and generate test scenarios.”
The test scenario also includes a simulation of how other road users react to AV. Cruise’s system for this is called Non-Player Character (NPC) AI, usually a video game term, but in this context, all cars and pedestrians in the scene that represent complex multi-agent behavior. Point to.
“Therefore, Morpheus, Road to Sim, and NPC AI work together in this very thoughtful way to enable more robust testing for rare and difficult events,” Gandhi said. increase. “And it really gives us the confidence that we can solve rare problems for similar problems now and in the future.”
Generating synthetic data helps Cruise AV target specific use cases, Gandhi said. Perhaps there is no reason other than the Autopilot ADAS system digging into Tesla under federal supervision. Repeated collisions with emergency vehicles..
“Emergency vehicles are rare compared to other types of vehicles, but they need to be detected with very high accuracy, so we use a data generation pipeline to create millions of simulated images of ambulances. Fire engine When Police car“In our experience, targeted synthetic data is about 180 times faster and millions of dollars cheaper than collecting road data, and the right combination of synthetic and real data. You can increase the related data in the dataset by more than one digit. “
Two custom silicon chips developed in-house
On GM Investor Day in October, Cruise CEO Dan Ammann invested heavily in Origin’s computing power to reduce costs by 90% and achieve profitable expansion over the next four generations. I outlined the company’s plans to do so. At the time, Ammann mentioned Cruise’s intention to manufacture custom silicon in-house to reduce costs, but did not fully admit to building chips using that silicon. TechCrunch had that theory.. On Thursday, Origin program chief engineer Rajat Basu tested these theories.
“Our 4th generation computing platform is based on in-house custom silicon development,” says Basu. “This is for our applications only. It allows focus, increases processing power, and significantly reduces component costs and power consumption. Computing is an important system from a safety standpoint. It has built-in redundancy. It adds an AV system to it, which means it is processing up to 10 gigabits of data per second, which consumes a lot of power. Using an MLH chip It allows you to run complex machine learning pipelines more intensively, increasing energy. Efficient without sacrificing performance. “
Cruise’s AI team has developed two chips. Sensor processing chips handle the edge processing of various sensors such as cameras, radar, and acoustics. Designed as a dedicated neural network processor, the second chip supports and accelerates machine learning applications such as large multitasking models developed by AI teams. According to Basu, machine learning accelerator (MLA) chips are just the right size to accurately solve a particular class of neural nets and ML applications, and nothing more.
“This keeps performance at a very high level and avoids wasting energy on doing things that aren’t value-added,” says Bass. “It can be paired with multiple external hosts or run standalone. It supports a single Ethernet network up to 25G and has a total bandwidth of 400G. The MLA chip we are putting into mass production It’s just the beginning. We’ll continue to manufacture it. It’s even more powerful while reducing power consumption. “
One of the things Cruise revealed during the event was eco-friendly, including not only the AV technology needed for a successful scale-up, but also a remote assistance operator to verify AV decisions when encountering unknown scenarios. I’m also thinking about the whole system. , Customer service, the car you actually want to drive, customer support and incident response, etc. can be handled efficiently and easily.
“To truly cross the gap from R & D to beloved products, artificial intelligence RoboticsOliver Cameron, Vice President of Products at Cruise, said at the event. Self-driving car Alone is not enough and is only the first step in a long and long journey. To truly build and extend the competitive products that millions of people embrace in their daily lives, they need to build a number of differentiated features and tools on top of a secure autonomous driving platform. there is. It’s not clear how these features need to be implemented. Especially if the company is solving safety issues head-on. “
Cruise shows plans for “how” to achieve robotaxis
Source link Cruise shows plans for “how” to achieve robotaxis
The post Cruise shows plans for “how” to achieve robotaxis appeared first on Eminetra.