When a ball bounces into the road, humans realize that a child may come running right after it and they slow their car down just in case. The same is true of seeing a single deer cross the street; our knowledge of the fact that deer often travel in herds means that we know to slow down and watch for additional deer following behind.
But autonomous vehicle technology doesn’t have those past experiences and knowledge that inform these decisions we make every day. As we move toward safe Level 5 autonomous vehicles, this is one hurdle that we’ll need to overcome. Multiple companies are already working on incorporating human feedback into AI technology.
Current Safety Concerns with Autonomous Technology
Safety is top priority in developing the technology used to power autonomous vehicles but it also presents a challenge. While these vehicles can be equipped to perform tasks like recognizing obstacles, taking evasive action to avoid obstacles and adjusting their speed according to road conditions, it isn’t yet possible to prepare the technology for all of the potential situations it could encounter while on the road.
Human drivers make split-second decisions during every drive, and they rely on their knowledge to help make these decisions – knowledge that autonomous vehicle technology doesn’t necessarily have. For instance, humans can recognize that it’s safer to drive over a paper bag than to try to slow down or avoid it while on a highway in busy traffic. The same is true of being able to differentiate between police cars with their lights flashing and an ordinary car approaching from behind.
Autonomous vehicle technology has been pre-programmed to react to situations in certain ways. An autonomous car will see a paper bag as a solid obstacle and will take (unnecessary) evasive action to avoid it. Because the autonomous vehicle has been approached by countless cars, it will continue doing what it’s been trained to do – drive – without realizing that the police car behind it requires a different action.
Our reactions to these situations are based on knowledge and past experiences, but these elements haven’t been introduced to autonomous vehicle technology – until now.
How Human Feedback and Correction Makes AI Technology Smarter
A new technology model could make autonomous technology “smarter” by incorporating human feedback and correction. MIT and Microsoft researchers describe this model as using human input to uncover “blind spots” where technology cannot rely on human knowledge and experience to navigate the world.
Using this model, the technology goes through simulation training, equipping it with the basic knowledge needed to navigate highways, roads, and the challenges that come with road travel. A human observes how the system interacts in the real world and provides feedback when the system makes mistakes, such as not recognizing an emergency vehicle.
The feedback data is combined with the typical training data to produce a model that can identify the types of situations where the technology needs additional information in order to navigate the situations correctly, taking the actions that a human would make.
So far, this method has been applied to video games. Incorporating the model into the development of autonomous cars and robots is the next step.
Applying Human Feedback and Correction to Autonomous Vehicles
MIT and Microsoft researchers have outlined how human feedback and correction can be further applied to AI and autonomous vehicle technology. The model system described above would first be put through simulation training where it would map out each situation and the best action to take in response. Next, humans would identify errors in the system’s operation in the real world.
Humans could demonstrate appropriate actions by continuously manually controlling the car and allowing the system to recognize instances where its actions would derive from the human’s actions.
Humans could also correct the system by taking the wheel and manually making a correction only when the car’s action is inappropriate. When the car’s actions are appropriate, the human would remain a passenger and would not operate the controls.
After the system receives this feedback and correction, an algorithm can identify the probability of particular situations being a “blind spot” where the technology may respond incorrectly. This data then helps the technology to act more cautiously when deployed in the real world.
The Implications Human Feedback Has on Autonomous Vehicle Technology Human feedback and the social contract technology can help to make self-driving vehicles safer and better able to navigate the countless judgement calls that we make as drivers every day. This human feedback-refined technology could be applied to all autonomous technology, including robots and drones. While it may be impossible to ever completely eliminate these blind spots and ensure that technology will always respond in a completely appropriate way, companies like Microsoft, MIT and Mobileye are already working to find the most effective ways to implement human feedback and reasoning in autonomous technology.