Wayve learns safety instead of relying on checkers

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 12:50

   part, how we prove safety, how we think about data in different continents. but i want to touch briefly on the ai driver itself and the main message here is that the way that we've built the driving stack is with safety as an intrinsic property versus an extrinsic bolt on property lots of systems today are either hybrid or compound av or they're purely modular sorry ai not av where they have different subsystems that are responsible for different aspects of the task and one feature of these systems it turns out is there's lots of checking yeah the canonical example i'll give you and then i'll come to this example that's on the screen is um maps lots of systems have high definition maps that are really like checkers like they say well look we think that we've mapped this before there's a stop sign right here we really need to

2. 13:46

   check uh that there's a stop sign here and stop for it and there's all sorts of challenges you can get into with systems like that like what if the stop sign isn't there what if it moved what if there's a temporary stop sign what if the stop sign is connected to a bus that's moving and you see all of these different failure modes and the the reality is that the checker systems tend to be the ultimate bottleneck and limit of system performance it's not impossible there are some fantastic systems out there today that you can get in and we can go take a we can go take a ride right now you know hundreds of miles around the bay area on on systems like these that are just absolutely fantastic they were built um at great expense and they require great operational overhead to maintain um and uh like those systems but especially a way if we're working on making that

3. 14:34

   a much more sustainable um system so uh net net we try to learn intrinsically what is safe behavior What you see on the screen here is a representation, a learned representation of safety. You can think of this as like a future occupancy map. The red areas are the don't drive in areas and then sort of a gradient. And you might have seen representations like this before. Those familiar with AVs or robotics where it just sort of looks like an occupancy map, a future occupancy map.

4. 15:03

   But really what this representation is saying is predicting where is and isn't safe to drive based on you see this bus coming up on the right. So that's a rear bus coming up fast, and it's telling you in the future it's not going to be safe or it's going to be marginally less safe to do a lane change into this lane. So we use this system not as a bolt-on checker but as an explicit head, a task that we learn on how to drive safely, and we use this to guide our ultimate driving trajectories.

5. 15:34

   Now, we, of course, have all of the lower-level systems that you might expect from a car today that you need to do to be compliant. these active safety systems, automatic emergency braking, lane-keeping assist, emergency lane-keeping assist is a soup of TLA's, three-letter acronyms I won't bore you with, but the main AI driver learns safety as an intrinsic property. So here's a representation. All right.