A Robot the Size of a Paper Clip Just Learned to Backflip 10 Times in 11 Seconds

A robot smaller than a paper clip spent Monday flying gymnastic routines in a lab at MIT, stringing together ten backflips in eleven seconds without drifting more than five centimeters off course. It was one thread in a week that also produced a cheaper, more accurate chatbot, a spacecraft built to fly itself with no help from Earth, and a free giveaway of the software behind Google’s factory robots. India spent the same days closing gaps that had stood for decades: a weightlifter’s 28-year wait for an Asian Games medal, a sport that had never once produced an Indian medalist, and a space partnership renewed for six more years.

Ten backflips in eleven seconds

MIT engineers gave their flapping-wing microrobot a new AI flight controller, and the tiny machine got dramatically better at flying. The robot weighs less than a paper clip and flies on flapping wings driven by artificial muscles rather than spinning rotors.

Under the new controller, its top speed jumped by about 450 percent. In tests reported this week, it strung together ten consecutive somersaults in eleven seconds while staying within five centimeters of its intended flight path, holding that accuracy even as researchers blew wind across it.

Real insects fly this way, using flapping wings for sharp, fast turns that fixed-wing aircraft can’t match. The MIT team isn’t chasing backflips for their own sake. A machine this small and this agile could eventually slip into collapsed buildings or other tight, dangerous spaces a standard drone is simply too big to reach.

A chatbot that makes fewer mistakes, for less money

OpenAI released two new models on Tuesday, GPT-6 Sol and GPT-6 Luna, into ChatGPT, its Codex coding tool, and its API. Both arrived with steep price cuts.

GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens through the API, down from $4 and $20 for the previous Sol model. Luna, the smaller and cheaper of the two, fell to $0.10 and $0.50 per million tokens.

The price cut didn’t come with a quality trade-off. On OpenAI’s internal factuality tests, GPT-6 Sol made about half as many mistakes as its predecessor, closing in on the reliability of OpenAI’s most capable model at a fraction of the cost. Sol and Luna rolled out immediately to Plus, Pro, Business, Enterprise, and Edu users, with a version of Luna also reaching free ChatGPT users through the desktop app.

A spacecraft that takes no orders from Earth

AstroForge, a startup building machines to mine asteroids, said the AI model it built in-house will take autonomous control of its next spacecraft. The mission, called Autonomy-1, is set to launch in 2027 on Stoke Space’s Nova rocket.

The plan is unusual even by space industry standards. After Autonomy-1 separates from its launch vehicle, AstroForge says the spacecraft will receive zero commands from Earth and attempt every mission objective on its own, while also carrying a NASA Goddard heliophysics instrument along for the ride.

The AI system, called Solo, is a transformer model trained on data from roughly 2,500 spacecraft sensors, combined with traditional flight-control algorithms and models of individual subsystems like power and navigation. AstroForge plans to fly Solo in shadow mode first, letting it make recommendations that engineers can check against its own performance before trusting it with full authority.

The company built Solo after losing contact with an earlier probe, Odin, in 2025. Rather than spend an estimated $200 million on a bigger ground communications network, AstroForge chose to put more intelligence on board instead.

Google’s robot software, now free to copy

Alphabet’s robotics unit, Intrinsic, open-sourced a chunk of its industrial robotics platform on Tuesday at ROSCon 2026 in Toronto, the annual gathering for developers of the Robot Operating System.

The release, called Intrinsic Core, is available under the permissive Apache 2.0 license on GitHub. It includes the same real-time control software, motion planning, and calibration tools Intrinsic uses in its own factory deployments, packaged so other developers can build with them directly instead of coding a robot’s basic capabilities from scratch.

One notable piece is pose estimation built on Nvidia’s FoundationPose system, which lets a robot locate and pick up a part without needing it held in a rigid fixture first, a common bottleneck in factory automation. Intrinsic framed the release as an attempt to give the wider robotics field shared infrastructure to build on, the way ROS itself has served research robotics for two decades.

The weightlifter who waited 28 years

India’s Mirabai Chanu won silver in the women’s 49 kilogram weightlifting category at the Aichi-Nagoya Asian Games on Wednesday, lifting a total of 198 kilograms: 87 in the snatch and 111 in the clean and jerk.

It was the first Asian Games medal of Chanu’s career, the one podium that had eluded a lifter who already holds an Olympic silver, a Commonwealth Games gold, and a world championship medal. It was also the first Asian Games medal won by any Indian weightlifter since Karnam Malleswari’s silver at Bangkok in 1998, a 28-year gap.

North Korea’s Ri Song Gum took gold, breaking the world record on her second clean and jerk attempt. Chanu told reporters afterward that her dream had finally come true.

A sport India had never reached the podium in

Jay Meena won India’s first-ever Asian Games medal in soft tennis on Wednesday, a bronze in the men’s singles event at a sport that has been part of the Games since Hiroshima in 1994.

Meena nearly didn’t make the medal rounds at all. In his quarterfinal against the Philippines’ Sherwin Nuguit, he trailed 3-5 in the deciding game, four points from elimination, before winning four straight points to close it out 4-3. He then lost the bronze medal match 2-4 to Japan’s Takuya Kurosaka after a 45-minute contest.

Soft tennis, a variant of the sport played with a softer, hollow rubber ball, had gone 32 years on the Asian Games program without a single Indian medalist. Wednesday ended that.

India and Europe extend a space partnership neither side wants to lose

The Indian Space Research Organisation and the European Space Agency signed an exchange of letters in Paris on September 10, extending their cooperation agreement by five more years, through January 2032.

The two agencies have already worked together on India’s Aditya-L1 solar mission and ESA’s Proba-3, and are expected to collaborate on India’s upcoming Chandrayaan-5 lunar mission as well. The renewed agreement widens the scope of cooperation beyond its original focus, adding space weather research and human and robotic exploration, and formally recognizing the growing importance of keeping orbital space usable for everyone.

It’s a quieter kind of news than a medal or a model release, the sort that rarely makes a headline on its own. But agreements like this one are what let missions like Chandrayaan-5 actually happen on schedule.