Solve roundabouts, Chazman’s turn and navigate without GPS and map data


By Alex Jones

While unguarded left turns can be nerve-wracking for novice drivers, many FSD beta participants still feel scared when their car is confronted with such a turn.

Exposed left turns are one of the more difficult driving maneuvers for humans and AI, as they make your vehicle vulnerable when turning (rather than turning with traffic).

Tesla quickly recognized his AI’s difficulties in executing these turns safely and worked diligently to improve the AI’s ability to recognize oncoming hazards.

Solve Chazman’s intricate left turns

FSD Beta 10.12 has improved unprotected left turn issues, but there are still some issues.

Chuck Cook (@Chazman) has repeatedly tested the FSD Beta’s ability to make unprotected left turns on a busy track.

It’s a complicated curve that required crossing three lanes before reaching the median and then turning left.

Often it involves crossing the three lanes and waiting at the median before finding enough space to get into a lane.

It’s complicated, even for humans. However, Elon Musk said in a recent tweet that Beta 10.13 would “mostly” solve Chuck’s complicated left turns.

Chazman on FSD Beta 10.12.2

Smoother crossings

Elon added that 10.13 would “smooth out intersection control” and hopefully reduce vehicle hesitation and steering judder that can occur at busy intersections.

Go deep into roundabouts

In addition to unprotected left turns, roundabout navigation has become the focus of FSD engineers.

While popular abroad, roundabouts are rarely used in the United States. Although there are estimated to be more than 2600 roundabouts across the country, FSD Beta needs to be able to handle them well.

FSD Beta’s navigation around roundabouts was patchy, although it could be argued that multi-lane roundabouts pose a challenge even for experienced drivers.

With FSD Beta 10.13, Musk says that Tesla is “going deep into roundabouts” and that it should be noticeably better in this iteration of FSD Beta.

There are certainly some complicated roundabouts that can challenge even the most experienced drivers. Some roundabouts have multiple lanes and may require changing lanes within the roundabout, some may even include roundabouts entering or exiting the roundabout such as: B. the “Magic Roundabout” in Swindon, England.

Magic Roundabout in Swindon, England

Navigate without map data

Musk also hinted at another important improvement that could be in 10.13: FSD Beta’s ability to navigate roads without map data.

The vehicle will be able to navigate to a specific GPS point or pinned location (eg, country roads), Musk says.

Dead reckoning (navigation without GPS)

Additionally, Elon alluded to the fact that Tesla is also working on the AI’s ability to complete “dead reckoning” (navigation based only on “inertial measurements, wheel motion, and sight”).

He cited underground car parks as an example of where FSD would need this ability to navigate without GPS or map data.

The car can do this by using its last known GPS location and then determining its future location using just a compass, wheel movement and speed.

Release Date of FSD Beta 10.13

Elon said he expects FSD Beta 10.13 to be available in about two weeks.

Why is it always two weeks? Tesla can use an agile approach to software development. In Agile development, work is planned in “sprints” lasting one or more weeks, with two weeks being a common length.

It’s possible that Tesla will use two-week sprints when planning and developing upcoming features.

For drivers who have experienced the intense nature of FSD left turns at busy intersections, it looks like you won’t have to wait long for FSD Beta 10.13.

By Gabe Rodriguez Morrison

NHTSA upgraded its investigation on autopilot

Earlier this year, the National Highway Transportation Safety Administration (NHTSA) launched an investigation into Tesla’s “phantom braking” problem, which occurs when the vehicle suddenly slows down.

In February, NHTSA stated that “rapid deceleration can occur without warning, randomly, and often repeatedly in a single drive cycle.” The problem is of particular concern on freeways, where sudden braking could result in a rear-end collision.

Since then, NHTSA has officially upgraded its preliminary investigation to a “technical analysis” that will be conducted before the agency issues a recall. Regulators will assess 830,000 vehicles after several of the automaker’s cars collided with stopped emergency vehicles while on autopilot.

Tesla’s Autopilot feature is designed to help drivers navigate roads using a combination of cameras and artificial intelligence to recognize other vehicles, pedestrians, traffic lights and more. Tesla instructs drivers to pay attention to the road and keep their hands on the wheel while using Autopilot.

In the past four years, 16 autopilot-activated Teslas have crashed into parked first responder vehicles, resulting in 15 injuries and one death. According to forensic data, most drivers had their hands on the wheel prior to the impact, in accordance with Tesla’s instructions.

In that investigation, regulators will determine whether or not Tesla’s Autopilot feature “undermines the effectiveness of driver supervision.”

“The investigation will evaluate the technologies and methods used to monitor, support and enforce driver engagement in the dynamic driving task during autopilot operation,” the NHTSA said.

This investigation covers approximately 830,000 Tesla vehicles built after 2014. Tesla introduced its Autopilot feature that year and later switched to its own Autopilot hardware in 2016 after ending its partnership with MobileEye.

Tesla has yet to respond to the escalated investigation into its Autopilot feature.

2 days since announcement

By Alex Jones

Elon says more animal visualizations are to come

Lions, tigers and bears, oh my god! While Dorothy’s animal list isn’t included in Elon Musk’s latest tweet about animal detection by Tesla’s neural network, dogs, cats, and horses are included.

Regarding Tesla’s object recognition abilities, Tesla appears to be preparing to release an update that will include new visualizations of various animals that may roam the car’s path.

As for animal recognition, Teslas currently only visualize dogs, but it will often display a dog even if it’s a different animal.

This is due to a lack of training, where the car cannot yet distinguish between a dog, horse or cat. It can be easily remedied by training their NN by adding and categorizing photos of additional animals.

It looks like we’ll be getting new animal visualizations soon. The most obvious are animals that are regularly found near roads, such as deer, horses, possums, or other creatures.

While automotive competitors have made strides in electric vehicle development after years behind Tesla, there’s no denying that Tesla’s neural networks are cutting edge.

These neural networks give Tesla the ability to collect live data from over a million participating vehicles. They use these huge datasets of real-world data to train their AI algorithms to identify objects that may pose a hazard to drivers.

Tesla claims that “a full build of autopilot neural networks involves 48 networks that require 70,000 GPU hours to train.”

While users have noticed drastic improvements in the car’s ability to recognize different car types (as seen through improved visualizations in 2022.16), the likely addition of new animals is an interesting addition.

With FSD Beta 10.12, Tesla added numerous visualization updates. Will more animals be added in FSD Beta 10.13, which is expected in about two weeks?

While it’s too early to tell, it remains exciting to see Teslas’ objection detection capabilities continue to improve.


Comments are closed.