Cruise Robotaxi Plan: How Self-Driving Cars Will Become Reality

Cruise Details Autonomous Vehicle Technology, Implicitly Contrasting with Tesla

During a detailed presentation Thursday evening, the engineering team at Cruise refrained from directly referencing Tesla. However, the underlying message was readily apparent to attendees.

Cruise, the autonomous driving subsidiary of GM, unveiled a comprehensive technical and deployment strategy. This roadmap was designed to demonstrate the superior safety and scalability of its autonomous vehicles compared to human drivers, even those utilizing advanced driver-assistance systems.

Demonstrating Technological Advancement

The event served as both a showcase of Cruise’s proprietary technology and a broader advocacy for autonomous vehicles. Engineers and product leaders detailed various aspects of their work.

Presentations covered the utilization of simulations, the in-house development of specialized chips and hardware, and the design of both the user application and the vehicle itself.

Roadmap for Commercial Deployment

This “Under the Hood” event expanded upon statements made by CEO Dan Ammann during GM’s recent investor day. He previously outlined the company’s plans for launching a commercial robotaxi and delivery service.

The initial phase will involve retrofitting Chevy Bolt vehicles, with a subsequent expansion to a large fleet of purpose-built Origin AVs over the coming years.

Recent Approvals and Cost Reduction Strategies

Cruise has recently received approval in California to operate commercial delivery services. They are currently awaiting one final permit to begin charging for driverless ride-hailing services.

The company anticipates significant cost reductions, enabling rapid scaling and expansion of its operations.

Cruise believes these advancements will allow for widespread deployment of its autonomous vehicle technology.

Leveraging Simulation for Scalability and System Validation

Cruise is utilizing simulation technology not merely for safety validation, but also as a crucial component in expanding its operations to new urban environments, reducing the need for extensive real-world testing.

While city mapping remains a necessary step for entry into new locations, the company aims to minimize the ongoing need for remapping due to inevitable environmental alterations such as lane adjustments or road closures. Upon entering a new city, Cruise employs a technology known as WorldGen, designed for the precise, large-scale creation of entire cities, capturing even minute details and unique layouts, as explained by Sid Gandhi, Technical Strategy Lead of Simulation at Cruise. Essentially, WorldGen provides the foundational setting for future simulation exercises.

To guarantee the creation of highly realistic virtual worlds, Cruise considers factors like illumination at 24 distinct times of day and various weather patterns. This commitment extends to meticulously measuring light output from a diverse range of street lamps within San Francisco.

“The combination of a highly detailed environment with procedurally generated urban landscapes unlocks our ability to efficiently scale our business to new cities,” Gandhi stated.

Gandhi then detailed the “Road to Sim” process, which converts real-world events captured by autonomous vehicles (AVs) into editable simulation scenarios. This approach prevents performance regression by ensuring the AV is continually tested against previously encountered situations.

“Road to Sim integrates data from perception systems with insights gained from millions of miles of real-world driving to reconstruct a complete simulation environment from road data,” Gandhi explained. “Once the simulation is established, we can generate variations of the event, modifying attributes such as vehicle and pedestrian types. This provides a simple and remarkably effective method for developing test suites that accelerate AV development.”

For scenarios that haven't been observed in real-world conditions, Cruise utilizes Morpheus. This system generates simulations focused on specific map locations, employing machine learning to automatically define numerous parameters and create thousands of challenging and infrequent scenarios for AV testing.

“As we progress in addressing the ‘long tail’ of rare events, our reliance on real-world testing will diminish. Testing a rarely occurring event requires thousands of road miles, which is not a scalable solution,” Gandhi noted. “Therefore, we are developing technologies to efficiently explore extensive parameter spaces and generate relevant test scenarios.”

The test scenarios also encompass simulating the reactions of other road users to the AV. Cruise’s system for this is termed non-player character (NPC) AI, a concept borrowed from video game development, but here representing the complex behaviors of all vehicles and pedestrians within a simulated scene.

“Morpheus, Road to Sim, and NPC AI collaborate effectively to enable more robust testing of rare and complex events,” Gandhi said. “This approach provides confidence in our ability to resolve both current and future similar challenges.”

The generation of synthetic data allows the Cruise AV to focus on specific use cases, Gandhi explained, specifically mentioning the identification and interaction with emergency vehicles – a point seemingly directed at Tesla, whose Autopilot ADAS system has faced federal investigation due to repeated collisions with emergency responders.

“Emergency vehicles are infrequent compared to other vehicle types, yet they require extremely accurate detection. We leverage our data generation pipeline to create millions of simulation images featuring ambulances, fire trucks, and police cars,” Gandhi stated. “Our experience shows that targeted synthetic data is approximately 180 times faster and significantly cheaper than collecting real-world road data. Furthermore, a strategic blend of synthetic and real data can increase the relevant data within our datasets by an order of magnitude or more.”

In-House Development of Two Custom Silicon Chips

During a GM investor event last October, Dan Ammann, CEO of Cruise, detailed the company’s strategy to substantially increase the computational capabilities of the Origin vehicle. The goal is a 90% reduction in costs across four successive generations, facilitating profitable scalability. While Ammann indicated plans for internal custom silicon manufacturing to achieve cost savings, he didn’t explicitly confirm the creation of a chip. However, TechCrunch anticipated this development.

On Thursday, Rajat Basu, the chief engineer for the Origin program, confirmed these expectations. “The compute platform for our fourth generation will leverage our internally developed custom silicon,” Basu stated. “This is specifically engineered for our needs, enhancing processing power and efficiency while dramatically lowering component costs and energy usage.”

Compute is a fundamentally important system from a safety standpoint, and incorporates built-in redundancy. Considering the AV system processes up to 10 gigabits of data per second, significant power consumption is inevitable. Basu explained that their MLH chip enables more focused execution of complex machine learning pipelines, leading to improved energy efficiency without sacrificing performance.

Two Chips Developed by Cruise’s AI Team

Cruise’s artificial intelligence division has engineered two distinct chips. The first is a sensor processing chip, designed for edge processing of data from cameras, radar, and acoustic sensors. The second chip functions as a dedicated neural network processor, accelerating machine learning tasks, including the large, multitask models created by the AI team.

Basu clarified that the machine learning accelerator (MLA) chip is optimally sized for a specific category of neural network and machine learning applications, avoiding unnecessary complexity. “This maintains an exceptionally high level of performance and prevents wasted energy on non-essential operations,” he said.

The MLA chip offers versatile connectivity, supporting single Ethernet networks up to 25G with a combined bandwidth of 400G. It can function independently or in conjunction with external hosts. Basu emphasized that the current MLA chip entering volume production represents only the initial phase of development.

“We will persistently enhance its performance and reduce its power consumption over time,” Basu concluded. “Continued improvements are planned to further optimize this crucial component.”

• Sensor Processing Chip: Handles data from cameras, radar, and acoustics.
• Machine Learning Accelerator (MLA) Chip: Accelerates machine learning applications.

The Cruise Operational Framework

During its recent presentation, Cruise emphasized its comprehensive approach to autonomous vehicle (AV) deployment. This extends beyond the core technology itself to encompass a complete operational ecosystem.

This ecosystem includes crucial elements such as remote support personnel. These operators are designed to confirm the AV’s judgments when encountering unfamiliar situations.

Key Components of the Cruise Ecosystem

• Customer Service: Providing support and assistance to riders.
• Vehicle Design: Creating a comfortable and appealing ride experience.
• Mobile Application: A user-friendly platform for managing support requests and incident reporting.
• Remote Assistance: Human oversight for complex or unusual driving scenarios.

Oliver Cameron, Cruise’s Vice President of Product, articulated that transitioning from research and development to a widely-used product necessitates more than just advancements in artificial intelligence and robotics.

A secure self-driving vehicle, while essential, represents only the initial phase of a prolonged development process. Scaling a competitive product for widespread daily use demands a suite of unique features and tools built upon a robust self-driving base.

The implementation of these features isn’t straightforward, particularly for companies still focused on resolving fundamental safety challenges. Successful scaling requires careful consideration of the entire user experience.

Cruise recognizes that a holistic strategy, addressing both technological and operational aspects, is vital for achieving long-term success in the autonomous vehicle market.