Tesla Autopilot is now enabling the car to sense the space around it thanks to the development of its occupancy network. Tesla’s Autopilot Software Director, Ashok Eluswamy shared a detailed thread on Twitter about the recent workshop organized by the Autopilot team. He also shared the workshop on Twitter.
Presented at CVPR this year some recent work from the Tesla Autopilot team, specifically regarding “occupancy networks” – our approach to common obstacle detection and using it to enable sophisticated collision avoidance. The whole thing here: (1/12)
Ashok Eluswamy (@aelluswamy) 21 August 2022
In the video and Twitter thread, Ashok explained how Tesla developed the Occupancy Network to give the car a sense of its surroundings. Human beings have the ability to understand the things around them at any given time. Is that car going slow or fast on the road? Do I, a pedestrian, have enough time to cross the street before I get hit? What’s that in the middle of the road? What is this falling from the sky? I must get out of the way.
These reactions to scenarios and split-second decisions come naturally to humans. Tesla’s Autopilot team is working to program vehicles to do the same and that would save lives. Imagine that the car is able to detect its surroundings accurately while the driver is not even paying attention. One example is sudden unexpected acceleration (SUA). Ashok said that the autopilot prevents about 40 such accidents every day.
The workshop was held at this year’s Conference on Computer Vision and Pattern Recognition (CVPR) in New Orleans in June. Ashok explained that the team has developed an occupancy network that enables the car to estimate the volumetric occupancy of everything around it.
Ashok explained that there are many issues in specific approaches such as image-space segmentation or pixel-wise depth of free space. The solution to those issues is the Occupancy Network.
In other words, the occupancy network enables the car to see the space around it and determine whether it can drive to that location. For example, if a UFO suddenly crashes in front of you while driving, you will react quickly in the safest way possible. The Autopilot team is training the software to do just that.
Ashok described how Occupancy Networks used Neural Radiance Fields (NERF). “The occupancy representation of these networks allows for discrete rendering of images (based on neural radians field task). However, unlike normal NERFs, which are per scene, these occupancy nets generalize to all scenes.”
These predictions are already used to prevent a lot of collisions. For example, Autopilot prevents ~40 accidents/day where human drivers accidentally press the accelerator at 100% instead of the brakes. In the video the autopilot automatically applies the brakes, protecting the man’s legs (7/12) pic.twitter.com/XtMssPT9cM
Ashok Eluswamy (@aelluswamy) 21 August 2022
You can read Ashok full twitter thread here And you can watch his presentation here. We’re a month past Tesla’s AI Day and I’m sure Tesla will share more about the life-saving technology as well as the Optimus bot.
Dr. Know It All recently published a video about the new 10.69 update and shared his thoughts about Occupancy Network.
In a message on Twitter, he told me, “The beauty of Occupancy Networks is that the car doesn’t need to know what The things he sees, he just needs to know that they Huh There to avoid them!”
Note: Johna is a Tesla shareholder and supports its mission.
Your feedback is important. If you have any comments, concerns, or see a typo, you can email me here [email protected], you can also contact me on twitter @Johna Crider1
Tesla Autopilot now enables the car to sense the space around it
