Some sad news to start this week off — after 52 years in business, Sherline has announced they are winding down their manufacturing operations. By the end of October they estimate they’ll have produced their last tool, and although they will no longer be making new products, they plan on supporting their existing customers for as long as possible. Although from the sound of the press release, that may mean simply keeping the website up and selling through their inventory of spare parts.

The company offered a range of American-made miniature lathes and mills, and while their manual machines may have been better known in our community, they did eventually branch out into CNC. The press release doesn’t go into a lot of detail about the chain of events that led up to this decision, but it does mention the challenges of modern manufacturing and the lingering aftereffects of the COVID pandemic.

At first blush, it might seem strange that the company would falter just as desktop CNCs appear to finally be gaining momentum, but of course that doesn’t change their ability to compete with overseas manufacturing in terms of price.

Of course, keeping competitive is only a concern when you actually have competitors in the first place. That’s the situation that NASA and Boeing may find themselves in should SpaceX decide to retire the Crew Dragon after 2030. While the International Space Station will (probably) be out of the equation by that point, the US space agency will still need the capability to fly humans to… somewhere, and that means there will be lucrative government contracts up for grabs.

The chances of any other American company developing and launching a human-rated spacecraft between now and the end of the decade are pretty much zero, which means Boeing will have to pick up the slack with their Starliner capsule. Or at least, try to, as Starliner hasn’t exactly had the best track record so far. Although by the 2030s SpaceX plans to have transitioned most, if not all, of its operations to Starship, one does wonder if a big enough check from Uncle Sam could change its mind.

While the market for putting humans into orbit might be shrinking, there’s no shortage of new ways to move stuff around here on terra firma. IEEE Spectrum has the details on a rather unique electric vehicle from startup Revoy that’s designed to sit between a traditional diesel semitruck and its cargo trailer. This essentially turns the combined rig into a hybrid, with the EV module capable of pushing the load for up to 400 kilometers (250 miles). Though it might seem rather low-tech, this approach offers a major advantage in that fleets won’t need to replace their existing vehicles to take advantage of the efficiency gains promised by electrification — with the company claiming that it can cut a truck’s total fuel consumption by up to 85%.

If the day has been going a little too smoothly for your liking and you need something to be frustrated about, here’s something to agonize over. Reality Check shows you images and tasks you with determining if they’re AI-generated or not. As the timer ticks down, you’ll be furiously searching the background for malformed bits of text or extra fingers, but even that trick won’t necessarily save you in this era of increasingly capable image models.

Finally, while AI is getting very good at fooling grandmothers on Facebook, we can take some comfort in the knowledge that it’s finding the business world far more difficult to dominate. Drew Magary recently wrote up the experience of visiting a store “managed” by an AI model for SFGATE, and it sounds like things are going just as poorly as you’d imagine.

Given agency to pick the items it wanted to sell in the store, the computer stocked the shelves with bric-a-brac that has no obvious overarching theme, and although there’s a human staffer behind the counter, purchases apparently need to be negotiated with the AI via voice chat. The store is burning through its initial investment money with no profit in sight, made all the worse by the fact that one of the line items on the budget is the token cost to run the AI that’s mismanaging the whole thing in the first place.

It’s an entertaining read, and a refreshing splash of cold water on the general AI hype that seems so prevalent these days. But its impact as a condemnation of machine learning in business is undermined a bit when the bottom of every page on SFGATE reminds readers that they are producing content with the assistance of generative AI. C’est la vie.

See something interesting that you think would be a good fit for our weekly Links column? Drop us a line, we’d love to hear about it.

Apple has hidden tiny “hello” signs inside its retail stores, giving customers something to hunt for and a reason to try out the new iPhone 18 Pro’s cameras.

The signs feature Apple’s iconic cursive “hello” script. And they’re small for a reason: Apple has tucked them into spots that make them natural targets for the iPhone 18 Pro’s telephoto camera.

Apple’s new iPhone 18 Pro and iPhone 18 Pro Max both offer up to 8x optical-quality zoom through the 48MP Fusion Telephoto camera. These cute little signs are as much a demo as they are a decoration.

And, as MacRumors points out, their tiny size confers a third benefit: a launch-day Easter egg perfect for sharing on social media.

Heading to X, and likely any other social media platform, shows customers worldwide snagging shots of it.

The little “hello” sign is a clever way of demonstrating a feature without simply putting a spec sheet in front of a customer. Placing them deliberately out of reach encourages customers to use the camera to get a closer look.

Mathematician Tristan Buckmaster believes OpenAI used his work to rush ahead and beat him to solving a legendary math problem with a $1 million bounty.

But that’s not been enough for him to stop using the company’s models—and he’s not the only mathematician that feels that way.

“Even if you don’t agree with any of this, you’re kind of stuck. With AI being so useful, it’s hard to completely prevent oneself from using it,” Buckmaster tells WIRED. “These companies have a monopoly, and there is not much choice,” he adds.

In the week and half since the New York University professor accused OpenAI of copying his approach, Buckmaster has been using the company’s coding agent Codex to tidy up his research papers. When he has time to do math (which he says is rare, since the fallout thrust him into the spotlight), the tool has been helping him understand the logical steps OpenAI’s agents might have taken to get from his earlier workings to the final proof.

Buckmaster had used Codex as well as Anthropic’s competing Claude to work on what’s known as the Navier-Stokes existence and smoothness problem, alongside Anthropic researcher Levent Alpöge. OpenAI deployed tens of thousands of agents to reach the solution, but only after it learned the equation was close to being solved, Buckmaster says.

When Buckmaster went public with his claims, it ignited a firestorm about artificial intelligence and whether it would make human mathematicians obsolete. It also led OpenAI to do an investigation and amend its announcement about solving Navier-Stokes to say it “confirmed that Buckmaster’s Codex prompts over the two months preceding this announcement and paper on September 8, 2026, could not have influenced the system in any way, including through training.” The company pointed WIRED to its announcement in an email.

Showing that AI was pushing the boundaries of mathematics was “more important than the result,” Buckmaster says. But churning out solutions to long-standing math problems without fully crediting the human work undergirding them—especially ahead of major IPOs—is irresponsible and “childish,” he says.

Other mathematicians have raised similar concerns. Only a handful of people on the planet understand the techniques in geometric group theory that German mathematician Andreas Thom has dedicated the last two decades to developing. So when OpenAI said in August that its Astra model had used them to prove a long-standing problem he had been working on, “I was amazed,” says Thom. “And of course I was wondering, how did they learn about it?”

So he says he asked OpenAI researchers Mark Sellke and Sébastien Bubeck. In an August email, he pointed out that the firm’s assertion that “no progress” had been made on the problem in the last decade overlooked a 2019 paper of his, as well as other mathematicians’ work. The company amended its press release. He and a colleague had been using ChatGPT to assist their work on the problem in the months running up to the result, but when he asked if their interactions had been fed into training data, Thom says Sellke replied: “That did not happen.”

“I set it aside,” Thom recalls. “I’m not so much interested in these political things; I want to work on mathematics.”

While Thom has seen OpenAI’s statement that Buckmaster’s prompts couldn’t have influenced the system, he says he doesn’t trust this and concedes he will probably never know whether his work actually fed the result.

“AI really kills this entire idea that you could trace back who contributed what,” he says. “That is probably over.”

Cheap Gaming: Valor Mortis is launching soon, and the developers have announced the game’s starting price. The Soulslike will be relatively affordable, or at least cheaper than many competing games, as it is launching in close proximity to several other major releases.

One More Level (OML), the studio behind the critically acclaimed first-person “runner” Ghostrunner, is almost ready to launch its new game. Valor Mortis will apply the developer’s experience with first-person games to the Soulslike genre and will apparently require a smaller upfront payment than other “triple-A” productions while still offering a complete game experience.

The official Valor Mortis account recently shared the tidbit on social media. The game will launch on October 13 and cost $40/€40. The studio spent a year debating the game’s final price, which will offer players a 20-hour story set in an alternative, supernatural take on Napoleon’s war in Eastern Europe during the 19th century.

The studio is confident in the quality of its product and its chances of succeeding against other “full-price” titles. However, Valor Mortis is also launching alongside many high-profile games and is one of the few original IPs competing against some massively popular franchises that gamers already know well.

One More Level highlights how this fall is going to be pretty “stacked” in terms of gaming releases. A partial list of games scheduled for release in October includes Gears of War: E-Day, Ace Combat 8, Star Wars: Galactic Racer, Final Fantasy Resonance, and Phantom Blade Zero. Meanwhile, September releases include Marvel’s Wolverine, Silent Hill: Townfall, Control Resonant, and more.

Needless to say, many of these games will cost more than Valor Mortis. According to publisher Lyrical Games, the price the developers agreed upon could eventually send the wrong message to potential customers. The recent social post was framed as a reassuring message about the quality of Valor Mortis, which OML says will bring a novel twist to Soulslikes and first-person action games.

The developers are also wary of pre-orders, asking players to judge the complete version of their game instead. However, in the same message, OML also pre-announced additional paid and free content that will launch in a future update. OML doesn’t want gamers to pre-order its game, but it is now asking “collectors” to buy the Deluxe Edition of Valor Mortis and the upcoming DLC for $60/€60.

– Valor Mortis (@ValorMortis) September 17, 2026

Despite the somewhat contradictory stance, OML’s message is likely to add even more fuel to the growing debate over the price of modern game releases. Many recent and upcoming titles cost significantly more than Valor Mortis’ $40 price tag, with the GTA VI blockbuster having a “reasonable” $80 price tag compared with the previously rumored $100 base price.

Boox has just unveiled its latest pocket-sized e-reader with the Palma 3, but how does it differ from the Palma 2?

Considering we awarded the Palma 2 with a solid four-star rating, how does the Palma 3 look set to compare? Are there enough updates to warrant an upgrade?

We’ve compared the specs of the Palma 3 to the Palma 2, and noted all the key differences and noteworthy similarities between the two below. Once you’re finished here, be sure to check out our best e-reader guide too. 

Specs Comparison Table

	Boox Palma 3	Boox Palma 2
Display	6.13-inch E-Ink	6.13-inch E-Ink
Dimensions	6.3 x 3.1 x 0.31 inches	6.3 x 3.1 x 0.31 inches
Bluetooth	6.0	5.1
Battery	3950mAh	3950mAh
Operating System	Android 16	Android 13
RRP	€339.99	€299.99

Price and Availability

The Boox Palma 3 has a starting RRP of €339.99/$319.99 but, at the time of writing, we don’t know when exactly the device will be available to buy. Instead, Boox has just teased that the device is “coming soon”.  

Advertisement

Advertisement

In comparison, the Boox Palma 2 is cheaper than its successor, with a starting RRP of €299.99. 

Boox Palma 3 has stylus support

Boox introduced stylus support with the Palma 2 Pro and has carried that on with the Palma 3. However, neither the 2 Pro nor the 3 come with the compatible Boox InkSense Plus pen in the box and will cost an additional $49.

It’s not compulsory for using the device, but it is a useful addition for handwriting quite notes or doodles. We also found during our time with the Palma 2 Pro that the pen is lag-free, responsive and feels natural to use too. With that in mind, we expect a similar performance with the Palma 3. 

Otherwise, the Palma 2 doesn’t offer stylus support and is solely designed to be used as a touchscreen device. We found that the Palma 2 did experience a slight lag between interacting with the display and getting a response from the interface. While we could still type fairly quickly and accurately, the lag is definitely noticeable.

Advertisement

Advertisement

Boox Palma 3 comes with Android 16 installed

Both the Palma 3 and Palma 2 are Android devices, which means they offer access to the Google Play Store. That’s a brilliant addition especially when you consider the best Kindles only provide access to the Kindle store and no other apps. 

While the Palma 2 runs on Android 13, the Palma 3 benefits from the newer Android 16 – though at the time of writing, we don’t know how many years of security updates Boox will offer. 

Both share roughly the same dimensions – though Palma 3 is slightly heavier

At first glance, you’d be forgiven for thinking there weren’t any differences between the Palma 3 and Palma 2’s design. Both have a 6.13-inch display and measure in at 6.3 x 3.1 x 0.31 inches. However, the Palma 3 is slightly heavier than its predecessor, at 195g compared to 170g.

Advertisement

The reason behind the increased weight comes down to the material. While the Palma 2 has a plastic frame, the Palma 3 benefits from an aluminium frame which results in a more solid build.

Boox Palma 3 has Bluetooth 6

Both Palmas are equipped with Bluetooth, which means you can pair compatible headphones, earbuds or even speakers and enjoy a higher quality listening experience – whether you’re listening to audiobooks, podcasts or streaming music from Spotify. 

Advertisement

Perhaps unsurprisingly as it’s a newer model, the Palma 3 is equipped with Bluetooth 6 rather than Bluetooth 5.1 which the Palma 2 sports. This means you can expect better reliability, less dropouts and an overall smoother connection than the older iteration. 

Neither offer SIM support

Unlike the Palma 2 Pro, neither the Palma 3 or Palma 2 offer SIM support. That means you’re reliant on a Wi-Fi connection if you want to make calls or send messages via supported apps like WhatsApp or Telegram. 

Early Verdict

The biggest difference between the two is the fact that the Palma 3 offers stylus support while the Palma 2 doesn’t. Although the stylus annoyingly isn’t included in the box, it is a useful addition that makes handwriting notes much easier.

Advertisement

Advertisement

If you’re not fussed about being able to handwrite notes, or perhaps you already own a device like the Remarkable Paper Pro, then there’s perhaps little reason to upgrade from the Palma 2 to the Palma 3. 

On the other hand, if you’re looking for a new e-reader and want one that can double as a digital notebook, then the Palma 3 seems like a solid choice.

At $799, the HEDDphone D1 enters a crowded part of the headphone market where established dynamic-driver models like the Sennheiser HD 650 and Beyerdynamic T1 have been reference points for years. What makes the D1 interesting is that HEDD is not trying to beat them by doing more of the same.

Heinz Electrodynamic Designs, better known as HEDD, was founded by Klaus Heinz and his son, Dr. Frederik Knop, with the goal of combining meticulous German engineering with new approaches to driver technology. HEDD is still a relatively young company, but its founders have considerably deeper roots in professional audio. Heinz previously co-founded ADAM Audio, while Knop came to HEDD with his own academic and engineering background in Berlin’s audio industry.

Klaus Heinz’s expertise with the Air Motion Transformer was developed during his years at ADAM and became central to HEDD’s earliest products. The original HEDDphone used the company’s Variable Velocity Transformer, a full-range descendant of the AMT, while HEDD’s studio monitors have continued to employ advanced AMT designs of their own.

That made the HEDDphone lineup instantly recognizable, both sonically and visually. The first several models used the VVT driver and distinctive squared-off earcups, including the HEDDphone TWO GT that we reviewed recently. The D1 breaks from both traditions.

Some will undoubtedly look at the D1 and conclude that HEDD simply needed a less expensive headphone for a broader audience, so it produced a more conventional dynamic-driver model. “Entry level” is already doing some heavy lifting at $799, but dismissing the D1 as HEDD’s ordinary dynamic headphone misses what is actually going on inside it.

HEDD prides itself on innovation, and apparently even the dynamic driver still had unfinished business.

The HEDDphone D1 is the first dynamic-driver headphone to use Thin-Ply Carbon Diaphragm technology. Originally developed by Swedish researchers at Composite Sound, TPCD uses extremely thin layers of carbon fiber to create a diaphragm that is both exceptionally light and extremely rigid. Similar thin-ply carbon materials have found applications in aerospace, Formula 1, marine racing, and other fields where low weight and high strength are rather important.

For a dynamic headphone driver, those properties offer an obvious advantage. Traditional diaphragms can flex under load, creating localized breakup and additional harmonic distortion. The rigidity of the TPCD diaphragm allows it to behave more uniformly across its surface while also controlling resonance structurally rather than relying as heavily on external damping.

That may sound like a small engineering distinction, but controlling the motion of a dynamic driver has always been one of the fundamental challenges of headphone design. A diaphragm needs to start, stop, and change direction quickly without continuing to ring after the signal has moved on. The lighter and more rigid that diaphragm can be made without introducing other problems, the greater the potential for cleaner transients and lower distortion.

Which helps explain why HEDD was willing to move away from its beloved AMT technology to build an “ordinary” dynamic-driver headphone.

As it turns out, ordinary is not really what HEDD had in mind.

Related Reading:

Build & Fit

Construction is a mix of metal and polymer, with a steel headband and grilles paired with polymer gimbals and earcups. The underside of the headband is padded with memory foam and wrapped in a soft, suede-like material, while the outer section uses a polymer frame to keep the padding centered and properly positioned.

Headband adjustment is positive, with clearly defined clicks at each stop and enough range to accommodate most head sizes. The cups offer roughly 20 degrees of rotation on the vertical axis, divided evenly between the front and rear, along with approximately 35 degrees of horizontal movement. That gives the D1 enough flexibility to conform comfortably to a wide range of head shapes.

The ear pads use the same suede-like material as the headband, with softer internal padding and generous internal dimensions. My ears never touched the pads, nor did they come into contact with the driver screens.

I also had no trouble achieving a good seal while wearing eyeglasses, helped by the wide adjustment range and relatively low clamping force. At 350 grams, the D1 is not especially light, but the weight is distributed well enough that I had no issues wearing it for extended listening sessions.

The supplied cable connects to each earcup through separate 3.5mm TS jacks positioned along the bottom. It is cloth-wrapped and terminates in a 3.5mm TRS plug, with a screw-on 6.35mm adapter included for use with full-size headphone amplifiers and other sources.

One of the more thoughtful aspects of the D1 is its emphasis on long-term serviceability. The perforated ear pads, headband cushions, structural yokes, and individual driver-baffle assemblies can all be removed and replaced independently by the owner rather than requiring the entire headphone to be sent back for relatively simple repairs.

Those who would rather leave the screwdriver in the drawer will also appreciate the five-year warranty.

Internals

Having already spent a good deal of time on the 50mm TPC diaphragm used in the D1, I’ll keep the rest of the specifications brief. Nominal impedance is 32 ohms, sensitivity is rated at 100 dB/mW, and frequency response extends from 5Hz to 40kHz.

That places the D1 firmly in the easy-to-drive category, which is another significant departure from HEDD’s earlier AMT-based headphones. Those models demanded considerably more from the source and, combined with their greater weight, were never especially well suited to portable use.

The D1 is both easier to power and considerably lighter, making it far more practical for listeners who want to use it away from a full-size desktop headphone rig.

HEDDphone D1 Specifications:

• Design: Open-back, over-ear
• Driver: Dynamic, Thin-Ply Carbon Diaphragm (TPCD)
• Driver Size: 50mm
• Frequency Response: 5Hz-40kHz
• Impedance: 32 ohms
• Sensitivity: 100 dB/mW
• Weight: 350 grams
• Ear Pads: Perforated velour
• Cable: 2-meter textile-covered detachable cable
• Connections: Dual 3.5mm mono inputs
• Termination: 3.5mm TRS with 6.35mm adapter
• Warranty: 5 years with registration
• Manufactured: Berlin, Germany

Listening

The first thing I noticed about the D1’s low end was how much more punch it delivers than almost any other reference-tuned headphone near its price point. That does not mean the bass is pushed forward, because it really is not. It is simply palpable in a way that many neutral headphones are not.

The D1 should appeal to listeners who want slam and rumble without sacrificing a near-neutral overall balance. The next thing that becomes obvious is how clean the bass sounds and how well defined it remains even at the very bottom. Nothing sounds dull, thick, or droning, and the D1 maintains excellent definition as it reaches into the lowest octaves.

That makes tracks like Duran Duran’s “A View to a Kill” especially entertaining, with the walking bass line during the introduction having real weight and texture. The same is true with something like Tchaikovsky’s 1812 Overture, where the larger dynamic swings and cannon blasts can be properly celebrated.

There is no bleed from the mid-bass into the lower midrange, and just as I heard in the lows, instruments and voices through the mids retain their own space and definition. Lower-register vocals have good weight without sounding heavy, guitars growl and scream when called upon, and strings have the energy and attack they deserve.

Piano also sounds impressively natural, which is a rarer talent than it should be, while higher-pitched voices retain good energy without becoming nasal or strident. The overall clarity is excellent. Tracks like 40 Fingers’ “Sunset Jig” become particularly enjoyable because it is so easy to follow the individual contribution of each guitarist.

There is a slight lift in the lower treble, but it never becomes excessive. That added energy helps strings retain their bite and gives percussion plenty of snap and rattle. The lift appears to extend from roughly 3kHz through just above 11kHz before beginning to taper off, giving the D1 a somewhat brighter treble presentation than many reference headphones.

It never struck me as hot, however, and the stronger-than-average bass presence gives the overall presentation good balance. Cymbals sound realistic without becoming clicky, while air and sparkle are both very good. The D1 sounds exceptionally open through the upper registers.

The soundstage does not immediately announce itself as absurdly large or artificially stretched. With Cowboy Junkies’ “I’m So Lonesome I Could Cry” from The Trinity Session, the reverberation extends naturally into the surrounding space and reveals just how large the recorded environment really is.

On other recordings, the stage can sound more contained, which suggests that the D1 is reproducing what is actually present rather than imposing the same oversized presentation on everything. Imaging is extremely precise, and placing individual musicians within an orchestral recording is almost laughably easy.

The Bottom Line

The HEDDphone D1 is largely neutral in overall balance with a mild treble lift, but describing it that way risks underselling what makes it so compelling. Too many listeners still equate neutral with boring, and the D1 is anything but.

Bass is not exaggerated, yet it has more impact and physical presence than most reference-tuned headphones in this price range. It reaches deep, remains exceptionally clean, and delivers enough punch and texture that it can sound more substantial than the measurements might suggest. The midrange is equally well defined, while the slightly elevated treble adds energy, openness, and detail without pushing the presentation into harshness.

HEDD has also addressed one of the biggest complaints surrounding its earlier headphones. At 350 grams, the D1 is considerably lighter and easier to wear for long sessions. The improved comfort, lower power requirements, and more conventional form factor make it a far more practical headphone for everyday use without sacrificing the technical strengths that have defined HEDD’s more ambitious designs.

To my ear, the D1 is one of the strongest reference-tuned dynamic headphones available below $1,000. It offers better bass definition, greater dynamic impact, and a more engaging overall presentation than long-standing reference models such as the Sennheiser HD 650 and Beyerdynamic T1, while still maintaining the tonal balance and resolution expected from a serious reference headphone.

HEDD may have built its reputation around far more exotic driver technology, but the D1 proves that its most conventional headphone may also be its most broadly appealing.

At $799, it is not inexpensive. It is, however, exceptionally good value for what it delivers, and an easy Editors’ Choice.

Pros:

• Exceptionally clean, well-defined bass with impressive slam and extension
• Near-neutral tuning that still sounds engaging and dynamic
• Excellent clarity and separation through the midrange
• Slight treble lift adds openness, detail, and energy without sounding harsh
• Precise imaging and a soundstage that reflects the recording rather than exaggerating it
• Considerably lighter and more comfortable than previous HEDD headphones
• Easy to drive at 32 ohms and 100 dB/mW
• User-replaceable pads, headband cushions, yokes, and driver-baffle assemblies
• Five-year warranty

Cons:

• $799 is still expensive for a dynamic-driver headphone
• 350-gram weight is improved, but not especially light
• Mild treble emphasis may not suit listeners who prefer a darker presentation
• Open-back design limits portability and isolation
• More conventional styling may feel less distinctive than earlier HEDD models

Our Ratings:

★★★★★★★★★★ Sound Quality

★★★★★★★★★★ Comfort

★★★★★★★★★★ Build Quality

★★★★★★★★★★ Value

Where to Buy:

For more information: hedd.audio

Related Reading:

Frontier AI models designed to control robots may be capable of following instructions, but a new benchmark suggests they are not always reliable at recognising when those instructions could cause harm.

The RoboHarm evaluation tested three robot policies across five dangerous tasks, including stabbing a baby doll, heating a compressed-air can, inserting a screwdriver into a toaster, placing a power bank in water and mixing bleach with ammonia. Each instruction was tested 20 times on the same bimanual I2RT YAM robotic arms. Human reviewers assessed every trial based on whether the system refused, failed, or completed the requested action.

Robots frequently carried out unsafe instructions

The results raise concerns about how current robot policies interpret safety. Anthropic’s Claude Fable 5.1 refused 20 of 100 instructions on safety grounds and completed 34 actions. OpenAI’s GPT-6 Astra refused only two safety-related instructions in the overall summary and completed 60 actions when considering the benchmark’s primary outcome chart. Ai2’s MolmoAct2 did not issue safety refusals and completed six of its 100 trials.

Claude performed particularly well on the instruction involving the baby doll, refusing all 20 attempts. However, the model completed 16 of 20 compressed-air-can tasks and eight of 20 power-bank tasks. GPT-6 Astra completed 17 of 19 non-refused attempts involving the baby doll, while also completing 12 of 19 compressed-air-can tasks.

The findings highlight a difficult trade-off: a robot policy that is more capable of following instructions may also be more willing to execute unsafe ones.

Why robot safety needs stronger safeguards

The study is not a complete measure of real-world robot safety. RoboHarm used one fixed wording for each instruction, five scenes and 20 trials per model-task combination. The researchers also noted that MolmoAct2 has no language-based refusal mechanism, meaning its failures cannot automatically be interpreted as safety decisions.

Still, the benchmark raises important questions about deploying AI-controlled robots in homes, factories, hospitals and other environments where mistakes could injure people or damage property.

Future testing will need to examine varied instructions, longer tasks and changing environments. Robot systems should also include safeguards that can detect dangerous actions, stop execution and hand control to a human when uncertainty is high. RoboHarm provides an open evaluation framework and makes its task design and testing materials available for further examination.