Physical AI in automotive demands robust edge engineering

<p><b>Intelligence must be integrated early to satisfy expectations, balance costs and meet strict safety requirements.</b></p>

<p>Physical AI has become a fundamental technology in the automotive sector, playing a key role in transforming the industry.&nbsp;</p> <p>It is at the heart of advanced driving assistance systems (ADAS) and essential to the progression toward self-driving vehicles. Without the presence of physical AI, autonomous automobiles wouldn’t be able to make the quick decisions necessary to protect drivers and pedestrians in safety-critical situations.</p> <p>Physical AI is also crucial to automotive manufacturers as they seek to enhance the in-vehicle user experience (UX). Key technologies, such as voice-controlled assistants that allow drivers to hyper-personalize their in-car environment, depend on it. The arena of physical AI is proving to be a key battleground for brands as they fight to differentiate themselves in the new era of software-defined vehicles (SDVs).</p> <p>While AI is integral to future success, the global automotive manufacturers we work with typically face several deployment challenges. For example, when&nbsp; decisions are required in milliseconds, solutions cannot be dependent on cloud computing—the latency would pose an unacceptable safety risk. When vehicles move, they can also hit connectivity blackspots, which would critically delay actions and detrimentally impact UX.</p> <p>So how can manufacturers enable what we call “AI-defined vehicles” to ensure their brand remains competitive in this evolving market? Based on our work with leading players in the sector, we’ve developed several thoughts and recommendations.</p> <h4>Physical AI requires robust edge engineering</h4> <p>To embed intelligence within a vehicle, OEM and tier 1 manufacturers must first overcome numerous hard constraints. Physical AI solutions deployed in an automobile must be able to infer, understand and react in real time. This puts a heavy demand on compute capacity, memory, power consumption, thermal limits and form factors.</p> <p>It is possible to add GPUs to manage this, but that’s rarely viewed as a viable economic option. The alternative is robust <a rel="noopener noreferrer" target="_blank" href="/content/cognizant-dot-com/us/en/services/cognizant-platforms/neuro-edge-generative-ai.html">edge engineering</a>. This enables trade-offs to be made without losing the accuracy, responsiveness and safety levels required.</p> <p>The first step we usually take to achieve this is to reduce the demands being placed on each semiconductor platform. The second step is to architect systems that can handle the specific requirements and ensure the performance delivered is uncompromised.</p> <h4>Optimize models to reduce hardware demands</h4> <p>One way we reduce the demands placed on the system is by optimizing the AI model needed. In most cases it is not necessary to deploy a powerful large language model (LLM) capable of answering every conceivable question—just those relevant to the requirements of the task.</p> <p>Optimization techniques such as model compression and quantization allow us to lower computational demands. We can then bring technologies such as natural language interfaces into vehicles without overwhelming the selected hardware.</p> <p>We used this technique to develop a <a rel="noopener noreferrer" target="_blank" href="/content/dam/connectedassets/cognizant-global-marketing/marketing-channels/cognizant-dotcom/en_us/services/documents/physical-ai-engineering-intelligence-in-the-real-world.pdf">voice-controlled massage seat</a> in partnership with a global automotive supplier. Designed for truck drivers, this seat is improving safety and comfort on long journeys. The deployment of a natural language interface removed a potential source of distraction, as drivers don’t need to reach for knobs or buttons. We developed a solution with an optimized LLM that runs at the edge. It delivers a responsive in-vehicle experience that meets strict safety requirements.</p> <h4>Design systems to handle the workload</h4> <p>Hardware selection is also helping us create systems that can handle the workload generated by physical AI. Advances in lower power GPUs and AI accelerators are making it possible to support various functional requirements across multiple generations of vehicles in an economical way.</p> <p>If demands on the chipsets remain high, we can also design systems with a distributed edge architecture that allows us to offload specific tasks onto nearby devices, such as a mobile phone. This is a good solution when manufacturers are not in a position to commit to a full platform redesign.</p> <p>As we move into the SDV era, we are also working to simplify applications, which will further reduce the demand on semiconductors within vehicles. This is freeing up computational power and memory for other purposes.</p> <h4>Integrate intelligence early in the design and development phase</h4> <p>AI is integral to future automotive trends. It is therefore essential that intelligence is integrated early in the design and development phase of any solution. Careful orchestration of the software, hardware, data platforms and vehicle architecture is necessary to ensure an end-to-end data flow.</p> <p>Embedding intelligence in the design and development phase also allows for early validation of system performance. New-generation virtualization platforms enable us to test latency, safety, power consumption and operational resilience. These tests also assess the real-world impact before solutions are deployed in live environments. This provides manufacturers reassurance that new solutions will perform reliably and efficiently throughout a vehicle’s lifecycle.</p> <p>This engineering process allows automotive manufacturers to deliver scalable AI projects that can provide measurable improvements to the driving and in-vehicle experience. It also ensures brands are satisfying rising expectations among vehicle owners and creating a competitive advantage while balancing costs and compliance concerns.</p> <p><i>To learn how physical AI can be embedded in your automotive manufacturing processes, see our whitepaper,&nbsp;<a href="/content/dam/connectedassets/cognizant-global-marketing/marketing-channels/cognizant-dotcom/en_us/services/documents/physical-ai-engineering-intelligence-in-the-real-world.pdf">Physical AI: Engineering intelligence in the real world</a>.</i></p>