Connect with us

News

Google’s DeepMind unit develops AI that predicts 3D layouts from partial images

[Credit: Google DeepMind]

Published

on

Google’s DeepMind unit, the same division that created AlphaGo, an AI that outplayed the best Go player in the world, has created a neural network capable of rendering an accurate 3D environment from just a few still images, filling in the gaps with an AI form of perceptual intuition.

According to Google’s official DeepMind blog, the goal of its recent AI project is to make neural networks easier and simpler to train. Today’s most advanced AI-powered visual recognition systems are trained through the use of large datasets comprised of images that are human-annotated. This makes training a very tedious, lengthy, and expensive process, as every aspect of every object in each scene in the dataset has to be labeled by a person.

The DeepMind team’s new AI, dubbed the Generative Query Network (GQN) is designed to remove this dependency on human-annotated data, as the GQN is designed to infer a space’s three-dimensional layout and features despite being provided with only partial images of a space.

Similar to babies and animals, DeepMind’s GQN learns by making observations of the world around it. By doing so, DeepMind’s new AI learns about plausible scenes and their geometrical properties even without human labeling. The GQN is comprised of two parts — a representation network that produces a vector describing a scene and a generation network that “imagines” the scene from a previously unobserved viewpoint. So far, the results of DeepMind’s training for the AI have been encouraging, with the GQN being able to create representations of objects and rooms based on just a single image.

As noted by the DeepMind team, however, the training methods that have been used for the development of the GQN are still limited compared to traditional computer vision techniques. The AI creators, however, remain optimistic that as new sources of data become available and as improvements in hardware get introduced, the applications for the GQN framework could move over to higher-resolution images of real-world scenes. Ultimately, the DeepMind team believes that the GQN could be a useful system in technologies such as augmented reality and self-driving vehicles by giving them a form of perceptual intuition – extremely desirable for companies focused on autonomy, like Tesla.

Advertisement

Google DeepMind’s GQN AI in action. [Credit: Google DeepMind]

In a talk at Train AI 2018 last May, Tesla’s head of AI Andrej Karpathy discussed the challenges involved in training the company’s Autopilot system. Tesla trains Autopilot by feeding the system with massive data sets from the company’s fleet of vehicles. This data is collected through means such as Shadow Mode, which allows the company to gather statistical data to show false positives and false negatives of Autopilot software.

During his talk, Karpathy discussed how features such as blinker detection become challenging for Tesla’s neural network to learn, considering that vehicles on the road have their turn signals off most of the time and blinkers have a high variability from one car brand to another. Karpathy also discussed how Tesla has transitioned a huge portion of its AI team to labeling roles, doing the human annotation that Google DeepMind explicitly wants to avoid with the GQN. 

Musk also mentioned that its upcoming all-electric supercar — the next-generation Tesla Roadster — would feature an “Augmented Mode” that would enhance drivers’ capability to operate the high-performance vehicle. With Tesla’s flagship supercar seemingly set on embracing AR technology, the emergence of new techniques for training AI such as Google DeepMind’s GQN would be a perfect fit for the next generation of vehicles about to enter the automotive market.

Simon is an experienced automotive reporter with a passion for electric cars and clean energy. Fascinated by the world envisioned by Elon Musk, he hopes to make it to Mars (at least as a tourist) someday. For stories or tips--or even to just say a simple hello--send a message to his email, simon@teslarati.com or his handle on X, @ResidentSponge.

Advertisement
Comments

Elon Musk

Elon Musk shares unbelievable Starship Flight 10 landing feat

Flight 10’s Starship upper stage demonstrated impressive accuracy when it came to its target landing zone.

Published

on

Credit: SpaceX/X

SpaceX CEO Elon Musk recently shared an insane feat accomplished by Starship’s upper stage during its tenth test flight.

Despite the challenges it faced during its return trip to Earth, Flight 10’s Starship upper stage demonstrated impressive accuracy when it came to its target landing zone.

Against the odds

Musk’s update was shared on social media platform X. In a conversation about Starship upper stage’s return to Earth, Musk revealed that the upper stage splashed down just 3 meters (under 10 feet) from its intended target. Considering the size of the Starship upper stage and the ocean itself, achieving this accuracy was nothing short of insane.

Starship Flight 10 was a success as both the Super Heavy booster and Ship upper stage completed all their mission objectives. However, videos and images released by SpaceX showed the upper stage’s heat shield scorched golden-brown and parts of its aft skirt visibly missing. The flaps and other surfaces also bore signs of heavy stress from reentry.

SpaceX highlighted this in a post on X: “Starship made it through reentry with intentionally missing tiles, completed maneuvers to intentionally stress its flaps, had visible damage to its aft skirt and flaps, and still executed a flip and landing burn that placed it approximately 3 meters from its targeted splashdown point,” SpaceX noted.

Advertisement

A key milestone

The result stands in stark contrast to Starship’s earlier test flights this year, when all three prior upper-stage flights in 2025 ended in premature breakup before splashdown. Flight 10 not only marked the first successful splashdown of the year for the Starship upper stage, but it also delivered near-perfect precision despite its battered state, according to a Space.com report.

For SpaceX, this success is a critical proof point in developing a fully reusable launch system. A spacecraft capable of surviving severe reentry conditions and still landing within meters of its target underscores the robustness needed for future missions, including orbital payload deliveries and, eventually, landings on the Moon and Mars.

Continue Reading

News

New Tesla Model Y Performance launches from Giga Berlin

The vehicle is produced at Gigafactory Berlin and is available to order now in Europe and the Middle East.

Published

on

Credit: Tesla

It took some time, but the new Tesla Model Y Performance is finally here. The new Model Y Performance features a blend of aerodynamic improvements, upgraded interior comforts, and high range enabled by new battery cells. 

The updated Model Y Performance is produced at Gigafactory Berlin and is available to order now in Europe and the Middle East. First deliveries are expected in 1-2 months.

Key Model Y Improvements

The new Model Y Performance sharpens the vehicle’s design and driving dynamics while adding subtle interior refinements. The revised variant now delivers 0–60 mph in 3.3 seconds, slightly quicker than its predecessor’s 3.5 seconds, while offering an EPA-estimated 308 miles of range, just about 1% less than the non-performance Dual Motor All Wheel Drive variant. Top speed is listed at 155 mph.

The exterior of the new Model Y Performance features new front and rear fascias, along with a carbon fiber spoiler designed for greater downforce and reduced drag. Staggered wheels and tires provide improved steering precision and grip, while high-performance brakes offer enhanced pedal feel and better heat management. The ride is supported by adaptive suspension that adjusts damping based on road conditions, paired with unique drive modes tailored for high-speed performance. Ground clearance is listed at 6.1 inches, and weight is listed at 2,033 kilograms (4,482 pounds).

Tech and Interior

Inside, Tesla has added carbon fiber decor, expanded ambient lighting in the footwells and door pockets, and upgraded seating. The first-row sport seats now include power recline, power tilt, heating, ventilation, and powered thigh extensions for added support during cornering. Rear passengers receive perforated heated seats with power recline. A new 16-inch QHD center touchscreen anchors the cabin’s technology suite.

Advertisement

Additional upgrades include eight exterior cameras, with the refreshed design introducing a new forward-facing unit. The high-density battery pack also boosts charge capacity but also helps maintain range despite the Performance model’s added power output.

Continue Reading

Elon Musk

Elon Musk reveals when SpaceX will perform first-ever Starship catch

“Starship catch is probably flight 13 to 15, depending on how well V3 flights go,” Musk said.

Published

on

Credit: SpaceX

Elon Musk revealed when SpaceX would perform the first-ever catch attempt of Starship, its massive rocket that will one day take life to other planets.

On Tuesday, Starship aced its tenth test flight as SpaceX was able to complete each of its mission objectives, including a splashdown of the Super Heavy Booster in the Gulf, the deployment of eight Starlink simulators, and another splashdown of the ship in the Indian Ocean.

It was the first launch that featured a payload deployment:

SpaceX Starship Flight 10 was so successful, it’s breaking the anti-Musk narrative

SpaceX was transparent that it would not attempt to catch the Super Heavy Booster, something it has done on three previous occasions: Flight 5 on October 13, 2024, Flight 7 on January 16, and Flight 8 on March 6.

This time, it was not attempting to do so. However, there are bigger plans for the future, and Musk detailed them in a recent post on X, where he discussed SpaceX’s plans to catch Starship, which would be a monumental accomplishment.

Musk said the most likely opportunities for SpaceX to catch Starship itself would be Flight 13, Flight 14, and Flight 15, but it depends on “how well the V3 flights go.”

The Starship launched with Flight 10 was a V2, which is the same size as the subsequent V3 rocket but has a smaller payload-to-orbit rating and is less powerful in terms of initial thrust and booster thrust. Musk said there is only one more V2 rocket left to launch.

V3 will be the version flown through 2026, as V4, which will be the most capable Starship build SpaceX manufactures, is likely to be the first company ship to carry humans to space.

Musk said that SpaceX planned to “hopefully” attempt a catch of Starship in 2025. However, it appears that this will likely be pushed back to 2026 due to timing.

SpaceX will take Starship catch one step further very soon, Elon Musk confirms

SpaceX would need to launch the 11th and 12th test flights by the end of the year in order to get to Musk’s expected first catch attempt of Flight 13. It’s not unheard of, but the company will need to accelerate its launch rate as it has only had three test flights this year.

Continue Reading

Trending