Proceedings of the 2021 5th CAA International Conference on Vehicular Control and Intelligence (CVCI), Tianjin, China, October 29-31, 2021 An Overview of the Zoox Smart Automatic Thermal Control Design -- An Autonomous Electric Taxi* 2021 5th CAA International Conference on Vehicular Control and Intelligence (CVCI) | 978-1-6654-0847-9/21/$31.00 ©2021 IEEE | DOI: 10.1109/CVCI54083.2021.9661241 Qinling Zheng and Paul Mueller Abstract— Proper thermal system control design can lead to enhanced vehicle battery economy, range, reliability, longevity, passenger comfort, and safety in an electric vehicle; advancements in thermal control design remain key as “new technologies, consumer demand, societal concerns, and government regulations emerge and evolve” [1]. This paper is an overview of the thermal control design used by autonomous, electric robotaxi start-up company -- Zoox. Different than many other traditional vehicle thermal control designs, the entire thermal system in the Zoox design features intelligence, robustness, energy efficiency, safety, and high performance to meet Zoox’s spirit of a rider focused taxi. Discussion of the challenges and solutions together with the high-level control design strategies are presented in the paper. each bench seat. The vehicle is designed with a tall roof and low floor providing comfort for taller riders and making it easier to enter and exit through the side doors. I. INTRODUCTION Zoox has been focusing on creating an entirely new autonomous, electric vehicle targeting the robotaxi market [2]. The company's approach is centered around the fact that a retrofitted vehicle is not optimized for autonomy. Zoox has applied the latest techniques in automotive, robotics, and renewable energy to build a symmetrical, bi-directional, battery-electric vehicle to solve the unique challenges of autonomous mobility [3-4]. Founded in 2014, the company built its first prototype in 2015, first demonstrated SAE Level3 autonomy in 2017, began durability testing of production prototypes in 2019, and established a production line in Fremont, California in 2020. The company’s name, Zoox, is short for zooxanthellae -- a mobile, algae-like organism powered by photosynthesis that thrives in mutually beneficial relationships with coral. To achieve the goal of making personal transportation safer, cleaner, and more enjoyable for everyone, the whole team at Zoox has created an entirely new form of transportation. Zoox will provide mobility-as-a-service in dense urban environments. Zoox will handle the driving, charging, maintenance, and upgrades for its fleet of vehicles, offering an effortless and enjoyable travel experience for its passengers. To differentiate itself from other existing autonomous vehicle designs, Zoox has spent the last few years working on outfitting its autonomous vehicle with the ability to drive both forward and backward, or “bi-directionally.” Zoox’s vehicle has also been designed to operate at speeds up to 75 mph. Fourwheel steering gives the robotaxi a tight turning radius of 28.2 feet making simple work of some U-turns and parking spaces that require other vehicles to use their reverse gear. Figure 1. Zoox Autonomous Robotaxi First Look [5] Artificial intelligence helps identify all objects perceived by the sensor suite, which when combined with GPS data and detailed map information provide the Zoox vehicle with the information it needs to determine a route and formulate specific routing around obstacles, congestion, construction, etc. Like most early autonomous vehicles, Zoox’s robotaxi is decked out in safety technology. There’s a crown of six rotating LIDAR pucks on the roof, each providing an overlapping 270-degree field of view that when combined with a full suite of radar and camera gear provides 360-degree awareness up to 150 meters in all directions virtually eliminating blind spots as well as providing redundancy in the event that a sensor fails. The interior design of Zoox’s vehicle is neutral and passenger friendly, materials used are premium as indicated by Figure 2. The cozy bench seats, which face inward (hence the term “carriage-style”) are surrounded by what looks like textured fabric. The seats also conceal what Zoox proposes is a radical rethink of how airbags work. There are cupholders and wireless charging mats between seats. And the ceiling has a starry sky pattern, the kind commonly seen in luxury vehicles. A small touchscreen at each seat is the most obvious tech found inside which can allow passengers to control music, radio, air conditioning, or see their route and estimated time of arrival. By designing the vehicle with inward facing bench seats, Zoox is able to provide a considerable amount of personal space for each passenger as shown in Figure 1. In addition, this allows the batteries and drivetrains to fit entirely underneath *Research supported by Zoox Inc. Q. Zheng is with the Zoox Inc. Foster City, CA 94404 USA (e-mail: jzheng@zoox.com). P. Mueller is with the Zoox Inc. Foster City, CA 94404 USA (e-mail: paulmueller@zoox.com). 978-1-6654-0847-9/21/$31.00 ©2021 IEEE Authorized licensed use limited to: Herricks High School. Downloaded on August 15,2023 at 04:58:31 UTC from IEEE Xplore. Restrictions apply. cannot be decoupled into smaller subsystems to enable singleinput-single-output (SISO) control system design. Figure 2. Zoox Interial Design [5] The front and rear battery packs combined store an impressive 133 kWh of energy, which is a little bigger than the packs that currently power Tesla’s most capable vehicles. These battery packs will last for 16 hours of continuous use and allow the robotaxi to be be recharged overnight and spend two-thirds of its day in use. Currently, Zoox is focused on testing on private and public roads as the company moves towards launching the first Zoox ride-hailing service. Since the vehicle is being designed for ride-hailing in urban environments, most of the fleet is testing in three cities: Las Vegas, Nevada; Foster City, California; and San Francisco, California. The company plans to launch an app-based ridesharing service. Its first target markets will be San Francisco and Las Vegas. After this brief review of the Zoox vehicle, one can imagine that such a rider focused fully autonomous robotaxi concept poses a great challenge with thermal control design. This paper will review the existing and some on-going thermal control designs in the Zoox vehicle. Control design challenges and a high-level design overlook will be presented in the following section. II. THERMAL CONTROL Thermal control in the Zoox vehicle is currently focused on, but not limited to, heating, ventilation, air conditioning (HVAC) control, and cabin climate control. HVAC control involves coolant loop and refrigerant loop control, while cabin climate control targets the regulation of the air loop in the cabin. When compared with traditional vehicle thermal design, the artificial intelligence and electric vehicle components (e.g., super-computing/memory components, high-voltage batteries, electric motors, related converters and inverters) need to be considered, since the Zoox robotaxi is an autonomous vehicle as well as an electric vehicle. It is a very complicated multiinput-multi-output (MIMO) system, with very tight inter-loop dynamic coupling among all three loops as shown in Figure 3. There are about one hundred sensors (such as temperature sensors, pressure sensors, valve position sensors, flowrate sensors) in the HVAC system and the coolant loop. There are more than a hundred input and outputs (I/Os) in the cabin climate control loop such as temperature sensors, air flowrate sensors, relative humidity sensors, air quality sensors (AQS), smoke detectors, solar radiation sensors, speed sensors, and vent positions sensors. Another characteristic in each loop is that there are multiple actuators for control input and multiple targets to be regulated as planned output. Furthermore, the input and output are not balanced, as they do not have 1-to-1 relationship, and therefore Figure 3. Inter-loops coupled dynamics Let’s take a close look at each loop. In the coolant loop, the control input consists of the radiator fan speed, multiple pump speeds, flow rate valve position, and a coolant route valve position. The targets to be regulated include the temperatures of all critical components, such as the motor, battery, GPU and CPU, to ensure they can operate at a healthy level without degrading performance. Meanwhile, the coolant loop is coupled with a refrigerant loop via a liquid cooled condenser and chiller. In the refrigerant loop, the input to the plant includes the compressor speed, condenser speed, evaporator speed, reheat door position, and purge door position. Outputs of the HVAC module include the HVAC discharge temperature, airflow, refrigerant pressure, and temperature. The air passes through the ventilation duct to the cabin to regulate the cabin comfort level. Airflow out of the duct and the air temperature leaving the duct are the main inputs to the air-loop in the cabin. In the cabin climate control loop, the temperature levels, airflow rate, and relative humidity level from passenger prospect of view must be considered when regulating the targets. The environment air quality inside the cabin such as smoke level, CO2 level, NO level, etc. should also be considered. One may also have noticed in Figure 3, there is also a trade-off between components and cabin cooling/heating capacity when the total heat/cool capacity is limited in certain ambient temperature cases. Therefore, optimized solutions are required to maximize the performance of the entire thermal system. Another challenge posed to thermal control engineers is that much of the dynamics in these three loops are characterized with nonlinearities and time delay which greatly decrease the stability margin especially phase margin. This is the main limitation to applying higher controller bandwidth, limiting the transient response time. Therefore, a significant number of lookup tables, which need to be carefully calibrated, are used to offer a fast pre-act action to compensate for phase lag. Robustness requirements are always important in vehicle control design. Thermal control is also required to be able to handle cases such as various weather conditions, different travel speeds, and different computing stress levels. Massive amounts of simulation tests, vehicle level validations, issue identifications, design improvement, as well as simulated and vehicle level shakedown tests have been completed and will continue as Zoox moves forward. The highest priority in vehicle design is safety and the same has been applied to the thermal system design. The Zoox thermal system is equipped with high accuracy AQS, allowing Authorized licensed use limited to: Herricks High School. Downloaded on August 15,2023 at 04:58:31 UTC from IEEE Xplore. Restrictions apply. the vehicle to switch the air mode from fresh to recirculation mode whenever it detects a dangerous level of NO, NO2, CO, CxHy which might threaten passenger health. A smoke detector monitors smoke level in the vehicle to ensure the cabin environment is always comfortable, healthy, and friendly to children. Highly efficient cabin air filters were selected and installed in the HVAC modules which can effectively absorb odors and gases, remove small particles, fungal spores, bacteria, and even neutralize allergens. Equipped with hundreds of sensors and I/Os in the vehicle thermal system, any single sensor fault or an inefficiency of I/O communication will cause system to shut down, affecting the performance or functionality of critical components such as motors, inverters, batteries, or AI related computational equipment. A welldefined fault handling mechanism is developed by the Zoox thermal system team to ensure that passenger safety has the highest priority while simultaneously considering comfortable level. The current thermal control design ensures that these control design challenges have all been taken into consideration. The next mission is to achieve better power consumption, lower noise vibration performance where better component design and optimized control will be applied. Comfort is always a top concern and the Zoox thermal control team hopes to offer the best travel experience for all passengers. As an autonomous robotaxi, bringing intelligence to the vehicle design will be reflected in every single system in Zoox’s taxi, including the thermal system. In the following section, some important smart characteristics in the thermal system design will be introduced. III. SMART AUTOMATIC THERMAL CONTROL The Zoox thermal design adopts an advanced control design, adaptive estimator/predictor, together while using a combination of modern machine learning (ML) techniques making the thermal system smart. Out of the many cuttingedge designs in Zoox’s thermal system, this paper will introduce a smart power budget design, a smart cabin temperature estimator, and a smart cabin climate target recommendation system. ambient temperature. An air forecast is used to predict environmental air quality. Supplied with this information, the ML model will offer a thermal system power budget used by the cabin climate control, coolant loop control, and refrigerant loop control. Figure 4. Smart thermal power budget B. Smart cabin climate estimator The thermal system is complicated and widely emerged in the whole vehicle system and almost covers every dimension of the vehicle attribute design. Therefore, thermal engineers at Zoox are working hard to minimize the usage of sensors to ensure a sleek inertia design, while offering a conformable cabin climate environment to all passengers. Decoupled zonebased climate control powers the micro-climate control in the Zoox robotaxi allowing each occupant to have their own personal comfort level of temperature and airflow. Adaptive state space observers capture the thermal dynamics caused by HVAC actuators, solar impact, vehicle speed, door open condition, and even RH level, and are built to provide breath level, chest level, and foot level temperature and air-flow rate for each quadrant as shown in Figure 5 and Figure 6. A. Smart power budget for thermal system The smart power budget system uses a time-series ML approach to predict the most likely amount of power consumption even before the trip begins, which will be very informative for trip planning and battery management. The model development and application schematic are described in Figure 4. To have a better prediction result, information such as day of year and time of day are used to represent the season, weather, and solar impact. These will be collected as training features of the ML model. Traffic information collected from a smart map and distance for suburb/urban environments are also considered to reflect the thermal cost for different vehicle velocity levels. A large amount of data will be collected to prepare our model for high accuracy and with high generality. In the application stage, after the model is trained and validated, the main inputs to the model will include several features. A map and GPS are used to provide vehicle speed estimation, travel distance, and compute stress level prediction. A weather forecast estimates solar impact and Figure 5. Offline observer fitting Figure 5 shows the how the adaptive state observer is fitted with collected test data. Figure 6 shows how the observer provides the estimated information that cannot be measured directly in the cabin to the cabin climate controllers. Temperature and airflow are estimated online and adaptively to minimize the feedback signals’ phase delay. Furthermore, this type of design allows the controllers to pre-act on typical process systems with large time-delay and ensure accurate control. Authorized licensed use limited to: Herricks High School. Downloaded on August 15,2023 at 04:58:31 UTC from IEEE Xplore. Restrictions apply. Figure 6. Online observer estimating C. Smart Temperature and Airflow Target Settings Another smart feature is to offer smart cabin climate control target via ML and an online training approach such as incremental reinforcement learning. As shown in Figure 7, the pre-trained recommended ML model is fed with available public information such as weather condition, air quality, and vehicle speed over the course of the trip to offer a recommended target temperature and comfort air vent position as an initial target value for the cabin climate control. Since Zoox will serve passengers as a taxi and may use a mobile application for initiating service, passenger information such as age and gender can be used as input to the ML model to adjust the target preferences. If a passenger has taken a Zoox taxi before, his/her previous preferences can be considered by the recommendation system. Otherwise, whenever a passenger is increasing/decreasing the cabin climate target setting via the touch screen panel, signals for reward/punish will be sent to the recommendation system based on its decision. Combining the offline and online ML training strategy is adopted to minimize the memory usage and computation power to allow the recommendation system to automatically adjust itself to be more personal while achieving a rider focused, effortless cabin comfort control goal. part of the entire vehicle system, its complexity and importance are motivations to design it right and design it well. The thermal control team enhanced modern control design with machine learning and an adaptive scheme to make it truly intelligent and allow the system to become smarter via online learning and self-adaptation. Some of the strategies have been developed, implemented, and are currently in validating and improving performance stages. There is still plenty of room for further improvements to performance to make the product even better. The whole Zoox team will try its best to deliver a comfortable and safe robotaxi to take you along your next journey. ACKNOWLEDGMENT This research is supported and funded by Zoox, Inc. We would like to thank the engineers from the Aero/Thermal/SCM integration team and the thermal system team in Zoox who provided insight and expertise greatly assisting the research. REFERENCES [1] [2] [3] [4] [5] G. Marshall, C. Mahony, M. Rhodes, S. Daniewicz, N.Tsolas, S. Thompson. “Thermal Management of Vehicle Cabins, External Surfaces, and Onboard Electronics: An Overview”. Journal of Engineering, Vol 5, Iss 5, October 2019, pp 954-969. "Zoox's driverless cars will operate like Lyft and Uber". CNBC. Retrieved 2019-02-11. Ohnsman, Alan. "Robotaxi Startup Zoox Becomes A Big Acquirer Of Tesla-Incubated Talent". Forbes. Retrieved 2019-02-11. "Zoox car: Mysterious Australian start-up worth $1.9b, but what do they do?". www.news.com.au. Retrieved 2019-02-11. https://zoox.com/ Figure 7. Smart temperature and airflow target prediction IV. CONCLUSION As the autonomous driving field becomes more and more prosperous, it also attracts interest from the fields of research and industry. Zoox, a young autonomous start-up company is dedicated to developing a rider-focused electric robotaxi. Given the many challenges on the road of exploring, the engineering teams never give up the attempt to offer Zoox’s future riders the most comfortable and safest experience while traveling in Zoox’s taxi. Although the thermal system is only Authorized licensed use limited to: Herricks High School. Downloaded on August 15,2023 at 04:58:31 UTC from IEEE Xplore. Restrictions apply.