Meet SOMAI, PA’s first Product — State of Mind AI
What is State of Mind AI (SOMAI)?
SOMAI is our proprietary flagship product, with a neural net model at its core, that was developed to replicate the intuition we instinctively have about other human beings. We rely on this human intuition every day to drive on our roads in a socially-acceptable manner. This human ability to understand the State of Mind of another human being evolved over thousands of years, and is a critical requirement for our customers’ L4 Autonomous Vehicles (AVs) to be accepted into society. We provide a simple C++ API that enables our customers to integrate SOMAI into their systems easily.
There are currently 2 key signals that SOMAI produces — an Awareness distribution and an Intent distribution. The former represents the level of awareness a pedestrian has of the ego vehicle, while the latter represents the level of intent a pedestrian has to imminently cross in front of the ego vehicle. To understand what this distribution represents — just imagine each AV has 1,000 people in the vehicle voting on what’s on the mind of each road user, the distribution of that voting is what SOMAI produces.
That sounds useful, but why wouldn’t your customers just build a ML model themselves to output that?
Yes they could certainly try (in fact some have). Part of our secret sauce lies in our unique ability to construct the appropriate annotation task and properly evaluate the resulting Machine Learning model, based on a unique approach from the field of psychophysics. Our founders and technical team are pioneers in this space from Harvard and MIT, and it has taken this group of experts many years to achieve the level of performance we think is necessary to be deployed on AVs.
Got it, so how does this help your customers?
Our goal is to help our customers achieve more socially-acceptable and understandable driving behavior. Any AV interaction with other road users where there is a difference between what a crowd of human drivers would do vs what an AV would do could be problematic. For example, if our human intuition tells us that it is safe to proceed past a pedestrian but an AV stops for her, not only is it inefficient, it could cause the AV to be rear-ended. On the other hand, if our human intuition tells us that we should stop for a pedestrian but an AV does not, it is at best annoying or impolite, at worst dangerous.
We at Perceptive Automata, along with our customers, are exploring many ways to use SOMAI to achieve that goal, including but not limited to :
- Predicting pedestrian trajectories
- Predicting road crossing behavior
- Predicting aggregate crowd behavior
- Adjusting minimum safe distances
- Assessing pedestrian / crosswalk risk levels
- Improving goal estimation
- Refining risk in occupancy maps
- As an input into a decision process (e.g. POMDP)
You can expect to hear more from us on these in the near future! In the meantime, feel free to get in touch if anything piqued your interest. You can email our VP of Business Development, James Gowers, at james@perceptiveautomata.com
