Can the APAS system be trained?

[Hiflyer808]

One thing that's not clear to me in the videos demonstrating APAS is whether the obstacle avoidance would be improved after multiple passes through the same set of obstacles. The official explanation of how it works talks about the drone using its cameras and sensors to do 3d mapping, so it seems it should get better over time. And with some of the demos when the drone just stops instead of going around an obstacle and a second attempt is made it does go around, but it's not clear if that's just because it started from farther away the second time.

I would really like to see some tests of difficult scenarios such as doing active track flying backwards in front of someone hiking on a forest path. Obviously, as we have seen, the first time it would just stop at a lot of obstacles and have to be steered away, but what if you went forwards and backwards through the same area several times? My hope is that eventually the drone would have fully mapped the area and you'd be able to get a really slick video where it smoothly avoided multiple trees flying backwards in front of you.

So, let's see those tests!

 

[Just_nick]

Their technology is nowhere near that level. None of the DJI drones do well with recognizing individual branches, and it's definitely not going to learn anything by repeating anything.

 

[Hiflyer808]

I don't expect the drone to learn to avoid obstacles it has trouble recognizing, like thin leafless branches. I'm talking about larger obstacles like tree trunks, where the drone does see it but stops instead of going around because there are other obstacles nearby so it doesn't know where to go. The hope is that after several passes through the area with assistance from the controller it would map out where all the obstacles were and figure out a smooth path through them. For best results the lowest altitude possible would be preferred so small branches would be less of a problem.

As I said before, there is some ambiguous evidence they do "learn", with a second attempt working when the first didn't. One simpler test of learning would be to do an attempt from too close so it stops, then an attempt from farther where it avoids successfully, then go back to the first distance and try and see if it still just stops.

 

[zwt100]

I’ve seen the marketing videos that allude to this as well so I’m interested if this is actually the case.

 

[thongbong]

One thing that's not clear to me in the videos demonstrating APAS is whether the obstacle avoidance would be improved after multiple passes through the same set of obstacles. The official explanation of how it works talks about the drone using its cameras and sensors to do 3d mapping, so it seems it should get better over time. And with some of the demos when the drone just stops instead of going around an obstacle and a second attempt is made it does go around, but it's not clear if that's just because it started from farther away the second time.

I would really like to see some tests of difficult scenarios such as doing active track flying backwards in front of someone hiking on a forest path. Obviously, as we have seen, the first time it would just stop at a lot of obstacles and have to be steered away, but what if you went forwards and backwards through the same area several times? My hope is that eventually the drone would have fully mapped the area and you'd be able to get a really slick video where it smoothly avoided multiple trees flying backwards in front of you.

So, let's see those tests!

Great question. I've wondered the exact same thing. A search on the internet pulls up basically no in depth details.

 

[AnzacJack]

The drone would need a heap of onboard memory to store mapped data and then a pretty decent processor to run all the calculations to link the mapped data to the drones position in space. I don’t think this is something far beyond current technology, however it would impact on size and power consumption.


You need this
The Skydio R1 might be the smartest consumer drone in the sky

 

[Hiflyer808]

The drone would need a heap of onboard memory to store mapped data and then a pretty decent processor to run all the calculations to link the mapped data to the drones position in space. I don’t think this is something far beyond current technology, however it would impact on size and power consumption.


You need this
The Skydio R1 might be the smartest consumer drone in the sky

Could be..DJI makes it sound like it creates a 3d map of the area, but yeah, it could be just on the fly, with no memory of the location of the objects just avoided. Still, these second tries when it avoids the object it stopped at the first time are intriguing. I suppose you're right, but let's see some systematic tests.

 

[DJI_dutch]

Great question. I've wondered the exact same thing. A search on the internet pulls up basically no in depth details.

It makes a 3D map of its surroundings. When you fly passed the same obstacle it might remember the map it has made of it. With its advanced algorithms it is a really smart quad!

 

[Mike4884]

Could be..DJI makes it sound like it creates a 3d map of the area, but yeah, it could be just on the fly, with no memory of the location of the objects just avoided. Still, these second tries when it avoids the object it stopped at the first time are intriguing. I suppose you're right, but let's see some systematic tests.

Yea seems like on the fly. And the farther away u are the more. Time it has to react to that specific obstacle...too close and it just stops

 

[scott rand]

I clipped a leaf less branch once with APAS on, luckily it didn't knock it down, just kept flying. Bigger objects, it does a phenomenal job. Always flies above or around people and tree trunks, and other larger objects.