Maybe if self-driving cars did not have to process all the details of every move they made, they might be able to more around more nimbly.
When we are driving on the road, our brains are processing so much information sub consciously. We are scanning the surrounding vehicles, anticipating their next move, and contemplating how we might respond. We even wonder how our moves might influence other drivers to act.
Robots have to do the same things that we do to operate in our world. Researchers have come up with new ways of enabling robots to better model our behavior. The researchers from Virginia Tech and Stanford University will be presenting their innovation to the annual International Conference on Robot Learning which takes place next week.
They plan to make robots more efficient by designing them to only analyze the major moves of other drivers on the road instead of analyzing each detail. This will enable them to predict their next actions faster and to respond swiftly.
A Theory of Mind
Robot makers are usually guided from Psychology’s theory of mind which posits that people are guided by their understanding of other people’s beliefs in their effort to empathize with and interact with them.
According to the Theory of Mind, young children master this skill and use it throughout their lives.
If robots can come up with a system for predicting the next moves of other actors, they can operate more efficiently on the road.
According to Stanford Professor Dorsa Sadigh, when two people are doing something simple like moving a table, they go by simple things like their discernment of the force from collaborators as they push or pull to make their next move.
A robot store could use this to remember basic descriptions of the actions of the agents surrounding. For example a robot playing a team sport could take note of the movements of its opponents and classify them as right, center, and left.
This data is enough to come up with two algorithms: one for predicting an opponent’s next move, and another for determining its own response.
The second algorithm will also track the opponent’s responses to the robot’s own responses, in order to allow the robot to learn how to influence its opponents.
The key feature of this innovation is that the robot handles minimal data and is therefore able to train itself as it goes along.
Usually, a robot in such a situation would remember not only the exact coordinates of every one of its opponent’s steps but also the direction in which they were moving.
This might seem like an approach that will yield greater accuracy, but that is not how the human mind operates. The human mind goes by simple clues and does not process too much information on the fly.
Future Uses
There are still a lot of questions that will only be answered in the long run. At the moment for example, researchers are working with the assumption that robots only interact in finite interactions.
Researchers assumed that robot cars involved in simulating self-driving were only experiencing single interactions with other cars in every training session. But that is not how driving works.
Cars interact with each other continuously and a self-driving car needs to keep learning and adapting its actions with every interaction.
Sadigh explains that the approach assumes that robot designers know how best to describe the behavior of other actors on the road. They thought of using ‘left’ ‘center’ and ‘right’ to describe the actions of opponents in an air hockey game.
These labels will become much less obvious when the interactions are not as simple. Researchers are still optimistic about their innovation being a game changer. By bridging the gap that separates human-AI interaction and multi-agent learning, scientist never stop learning because it is an important new field of research.