Delft, the Netherlands, January 14, 2025. Dewant Katare, a doctoral researcher at Delft University of Technology (TU Delft), presented an energy-aware adaptive framework for connected autonomous vehicles at the 2025 IEEE 15th Annual Computing and Communication Workshop and Conference (IEEE CCWC 2025), held January 6 to 8 in Las Vegas, Nevada. The framework reduces the energy that AI computation uses in the vehicle and on nearby edge servers by up to 40%, while model accuracy drops by less than 7%. Written alongside TU Delft researchers Marijn Janssen and Aaron Yi Ding, the paper examines an overlooked trade-off in autonomous driving: the perception models running on the onboard hardware pull an excessive amount of power from the battery, directly reducing driving range.
Why perception AI drains electric vehicle batteries
Autonomous driving and advanced driver assistance systems depend on perception pipelines, which generally consist of neural networks converting raw feeds from sensors such as cameras, radar, and lidar into a structured, interpretable map of the vehicle’s surroundings. Depending on hardware architecture and sensor count, running these models onboard demands anywhere from a few hundred watts to well over a kilowatt. In an electric vehicle, every watt pulled by computation drains the traction battery, lowering real-world driving range.
Most industry models are developed to achieve maximum accuracy. The usual answers to their power demand are larger onboard processors or offloading model serving tasks to the cloud. Katare targets software-level efficiency, replacing 32-bit math with lower and mixed-precision operations. The network extracts the same features from the sensor stream, but the compute hardware pulls far less energy.
How the framework saves energy
Katare’s framework applies approximate computing by running non-essential perception tasks at reduced precision, saving battery power wherever minor accuracy trade-offs do not impact decisions.
One part of the work targets Vision Transformers, image-recognition models that demand heavy computation. Katare designed software tools that combine variational inference with adaptive mixed precision, switching parts of a model between 8-bit, 16-bit, and 32-bit arithmetic depending on how much accuracy each part needs. In compute-heavy perception tasks, this approach saved up to 35% of energy.
Another part of the framework decides where each piece of computation runs. By pairing variational inference with lookup tables, it splits a model so that some layers run in the vehicle and others on nearby edge servers. Katare also built ARASEC, a runtime engine that allocates resources and partitions models across serverless edge-cloud setups. The framework is described in a paper published in IEEE Internet Computing.
“Real-world sensors never give a neural network perfectly clean data, so processing every input at full precision spends energy on noise,” said Dewant Katare. “Our results show that a vehicle can keep reliable perception while using much less of its battery for computation, which lets engineers treat efficiency as a design target from the very start.”
The research behind the framework
Katare carries out this work at TU Delft as an early-stage researcher in APROPOS, a €4 million Marie Skłodowska-Curie Innovative Training Network funded by the European Union’s Horizon 2020 programme. The network brings together international research groups working on energy-efficient computing. Katare’s first-author papers have appeared in IEEE Communications Surveys & Tutorials and the Journal of Systems Architecture. He has also published in the IEEE Open Journal of Intelligent Transportation Systems and at the ACM/SIGAPP Symposium on Applied Computing.
Katare came to academic research from industry. With a Master’s degree at Purdue University, he also worked on automotive processors and embedded systems at NXP Semiconductors and Endress+Hauser, where he saw deep vision models push hardware past its power limits. That experience brought him to the APROPOS network and to his research on energy-efficient edge AI, which centres on software methods for cutting the power that AI models consume.
The next stage of Katare’s work is to extend adaptive approximation from single vehicles to electric vehicle fleets and smart road infrastructure. His goal is to let AI models grow in capability without their energy usage growing at the same pace.
About Dewant Katare
Dewant Katare is an AI systems researcher at Delft University of Technology who works on approximate computing and energy-efficient edge AI for connected autonomous vehicles. He is an early-stage researcher in the EU Horizon 2020 APROPOS network and holds a Master’s degree from Purdue University.
Media contact
Dewant Katare
Email: [email protected]
Phone: +31 633578732
