Simulation is the Key to Industrializing AI in Automotive

Subscribe To Download This Insight

By James Hodgson | 3Q 2018 | IN-5251

Artificial Intelligence (AI) has played a key role in transitioning autonomous vehicles from abstract concept to the very core of most Original Equipment Manufacturer (OEM) strategies for the next decade. The widespread adoption of neural networks, running on high compute platforms–enabled OEMs and other autonomous vehicle developers, learn from experience, bypassing the need for the time-consuming and ultimately impossible task of anticipating and hardcoding for every possible scenario that the autonomous vehicle may need to navigate.

Registered users can unlock up to five pieces of premium content each month.

Log in or register to unlock this Insight.

 

Generating Experience

NEWS


Artificial Intelligence (AI) has played a key role in transitioning autonomous vehicles from abstract concept to the very core of most Original Equipment Manufacturer (OEM) strategies for the next decade. The widespread adoption of neural networks, running on high compute platforms–enabled OEMs and other autonomous vehicle developers, learn from experience, bypassing the need for the time-consuming and ultimately impossible task of anticipating and hardcoding for every possible scenario that the autonomous vehicle may need to navigate.

However, the adoption of the AI approach has had the side effect of requiring developers to accrue an exceptional number of driverless miles to validate an autonomous system. A 2016 study from the RAND Corporation determined that between 1 billion or even 100 billion miles would be necessary to demonstrate that an autonomous system was significantly safer than a manually controlled vehicle. Since the core marketing message for autonomous vehicles has been one of safety, OEMs are keen to have confidence in their autonomous vehicles before they are deployed.

Simulation Holds the Key

IMPACT


Accruing these miles physically would be an impossibility, given that the number of autonomous prototypes currently on the road numbers in the hundreds, rather than the thousands. It would take decades before OEMs could be confident that their autonomous system was better than the status quo. ABI Research calculates that if OEMs were to push ahead regardless, it could mean deploying 3 million vehicles over the course of 10 years before the vehicles collectively accrued 100 billion autonomous miles.

Clearly, there is a need for an alternative approach that allows OEMs to continue to take advantage of the deep-learning approach without the requirement to cover so much distance manually. Multiple vendors, including players from the semiconductor and autonomous software space as well as dedicated startups, have launched simulation solutions that can enable autonomous vehicles to be trained on virtual data. The accelerating effect this has on training is tangible. With 25,000 virtual vehicles, Waymo’s Carcraft simulation initiative can accrue around 8 million miles of driverless experience per day—as many as their physical prototypes have accrued since 2009.

The speed with which dedicated startups Cognata and Metamoto have gone from their initial founding to launching products and securing OEM customers demonstrates how powerful a tool simulation can be for autonomous system developers.

How to Deploy the Simulations

RECOMMENDATIONS


With the self-imposed deadlines for the deployment of autonomous vehicles only a few years away, OEMs are either acquiring technologies or developing partnerships to address the remaining points of friction which would impede autonomous vehicle deployment. Waymo, the only player with ambitions to deploy autonomous vehicles, has taken the step of developing a comprehensive in-house simulation, building on their existing leadership in physical prototype experience. For OEMs to effectively compete, they will be looking for partners who can help them to close the gap by fulfilling the following characteristics and approaches.

Simulation as a Service: With OEMs needing to close the gap with Waymo’s experience, service-based simulation offers the best way for OEMs to begin consuming simulations as soon as possible. Usage-based pricing, such as that adopted by Metamoto, can further reduce the barriers to entry.

Holistic Approach: Players such as NVIDIA and AIMotive have incorporated AI as part of their more holistic product offerings. NVIDIA, through their consistent Graphics Processing Unit (GPU) architecture in their Pegasus solution, allows their customers to train on their simulation networks with identical hardware to that which they will deploy on the road.

Software Model Integration: Software-in-the-Loop (SIL) approaches require the autonomous vehicle compute platform and its properties to be modeled in simulation. Indeed, simulation is only a useful exercise if the virtual autonomous vehicle behaves as the physical equivalent would in the field. Providing engineering support to OEMs to build robust and realistic software models of their autonomous vehicles can prove an important differentiator.

Sensor Partnerships: The process of building accurate sensor emulation requires a high degree of cooperation between the sensor technology developer and the simulation provider. The simulation provider will typically provide an emulated sensor feed for a simulated environment to the sensor technology developer, who will assess the quality of the emulation and then provide feedback. This circular process can continue through several iterations before an accurate emulation is achieved. Sensor vendors and simulation service providers can therefore develop a symbiotic relationship. Partnering with multiple sensor technology developers can ensure that the simulation is a more accurate reflection of reality. Meanwhile, sensor vendors can advertise the adaptable capabilities of their sensors with many potential customers at minimal cost.

Services

Companies Mentioned