Learn more about the questions on our minds when we think of the future of transportation systems around the world.
What technologies could be used to allow for better autonomous driving in the snow?
Computer vision is notoriously bad at detecting and recognizing objects in highly reflective or low contrast environments. Add to that the fact that nearly all training data in the past was on gravel or concrete roads, and demarcation lines are gone, as well as the fact that traction on the road is significantly reduced, and the majority of current self-driving systems become incapacitated. Even advanced sensors such as LiDAR scanning systems operate differently in snow and rain. How can we deploy self-driving systems in dynamic environments? How can we solve other edge cases such as driving on dirt roads?
What special sensors or systems are needed for special-use vehicles, such as garbage trucks, airport vehicles, and buses?
Varying in size, combined with the need for aligning with pickups and drop-offs, operating in smaller alleys or around pedestrians, may mean that special training and attention needs to be given to unique cases. We would love to hear about people working on solving these inevitable challenges.
What systems will allow airplanes to take-off and land more efficiently, closer together?
As air traffic control becomes more and more automated, as tracking between planes, ground crews and towers become more seamless and more frequent, autonomous systems can become better at increasing the utilization of our runways, reducing delays and increasing efficiency.
How can maintenance and diagnostics of vehicles [land, air, sea] be automated?
Sounds, heat, moisture, vibration, chemical and vapor sensors all can help machines understand their status better and better. As the quantity of this data increases by orders of magnitude, humans will struggle with interpretation in real-time. How can tools like Recurrent Neural Networks (RNN’s) help us machines understand time series data to build predictive models for failure, and integrate them into existing supply chain networks, to eliminate or minimize downtime?
How can we make data collection and labeling for HD 3D maps of every city, road, and alley more scalable?
Manual tagging and labeling of map scans and data are one of the most expensive parts of autonomous systems. How can we increase the efficacy and efficiency of mapping, and what other tools can we implement to increase the fidelity and reduce the costs for collecting HD map data? How can the maps be stored in a small enough size? How can retrieval become more seamless and efficient?
How can we integrate more datasets and user behavior models to reduce congestion?
Current traffic prediction and avoidance tools such as Google Maps and Waze are allowing us to use crowd-sourced data and road sensors to build sophisticated models for route optimization and management. Startups are attempting to integrate more and more data such as weather patterns, shopping behavior, and events to create more and more accurate models. We are searching for tools that can increase this even more.
How can we improve public transit through real-time routing, surge management, and autonomous vehicles?
Current routing for public transit is on fixed routes, with repeated patterns and limited capacity. How can we quadruple the capacity of our public transit systems while increasing comfort, coverage and reducing average wait times and travel times? Startups like Chariot are rethinking this using mobile infrastructure and dynamic routing to improve coverage. How can AI boost these efforts when we have hundreds of thousands of users on these platforms?
How can fueling become completely automated and operate efficiently, even staggered from surge timing?
How can robotic systems reduce the cumbersome process of charging or refueling, and be adaptable to any make or model of vehicle? How can this be made cheap and easy enough to work in homes or garages? How can this process be optimized to reduce load on the grid through solar and off-peak charging/scheduling?
How can we use advanced simulations to uncover edge cases necessary to train autonomous vehicles to be prepared when they happen?
One of the qualities of autonomous systems is their ability (as a network) to improve performance dramatically with time (cutting accident rate in half every 18 months, for example). Autonomous vehicles have faced barriers to adoption, regardless of being much less accident-prone than human drivers, but the “inexperience” of the system is still a large factor affecting regulation. What are strategies to accelerate this learning rate, preferably in a controlled/safe environment?
What is the right trade-off between on-board machine perception, and a well-documented, highly connected environment?
Autonomous vehicles have already started having very good success on well-paved and painted roadways in the US, but reliability could be further improved by creating a better database of video/images of vehicles, pedestrians, and infrastructure (like traffic signals, traffic signs, traffic monitoring cameras), and allowing these to communicate with each other. How much of the intelligence will be fully on-board the vehicle? How much information will be relied upon from non-vehicle sources?
What is the state-of-the-art security to prevent hijacking or digitally tampering with an autonomous vehicle, and what technical safeguards need to be added before mainstream adoption?
One of the major obstacles in consumers adopting new technology is viewing it as unproven in terms of security and privacy. What are some technical tools that can be added to help safeguard against any tampering of the vehicle or obstacle avoidance system?
What is the theoretical limit for V2V and V2I communication rates and connection accuracy, and what technology developments are necessary to approach this limit?
While vehicles may rely fully on on-board intelligence, they may be able to be made safer by allowing V2V and V2I to more quickly transmit information through a mesh-like network. What are some promising techniques for allowing vehicles to quickly connect when they are in proximity to each other, and securely send snippets of information about, for example, the state of the road ahead?
How will AI be used for new materials discovery and new manufacturing processes to make the product design and engineering in transit more efficient?
New materials discovery has already started to become mainstream in higher tech industries like defense. When will these technologies be applied to rubbers in wheels, for example, to create new materials that can more efficiently optimize cost, durability and material constraints?
What are some driver-focused sensors that can be deployed to decrease the accident rate and make cars safer for people to drive?
Emotion, attention, sobriety, and drowsiness are all emotions that might be detected through various passive sensors (vision, heart rate signature, etc). How can we incorporate this to the design of cars and trucks to reduce the frequency of accidents?