mailwal

What is Happening With Robot Hands?

Hardware FYI

What is Happening With Robot Hands?

Why robot hands are the bottleneck for humanoids, fusion-based alchemy, space mirrors that bring sunlight to the night, and more.

Topics worth your time: torque density and the gearbox tradeoff at the heart of dexterous hands, cross-sections of everything from the inside out, a fusion startup claims to have turned mercury into gold, and more.

Hardware FYI is brought to you with support from our partners.

Interesting Read: The Dexterity Deadlock

Robotics is currently the wild west of engineering, and it’s never been harder to separate meaningful signal from noise. The field isn’t new by any means (industrial robots were first installed on factory floors in 1959, albeit for narrow, specifically programmed tasks), but the rapid advances of the last few years have both startups and the largest companies in the world chasing something far more ambitious: general-purpose robots that work.

To understand the current bottleneck for helpful humanoids, look no further than the hand. It’s a single assembly that has to match roughly 27 DoF, deliver high grip force with enough compliance for delicate objects, sense what it touches, and be reliable enough to last years in service. By Tesla’s own account, a large part of the delay pushing Optimus deliveries to the end of 2027 is the hand’s design, which the company concedes is still in development as it scales to production; we’re building robots that do not have perfected hands.

For an engineering-first read of why hands are so difficult, The Dexterity Deadlock from Origami Roboticstraces much of it back to two factors: torque density, and the high-ratio gearboxes used to compensate for it.

Torque scales with r³ (Source: Origami Robotics)

A motor’s torque scales roughly with its volume, or the cube of its linear dimension. For example, shrink every dimension 10x and it’s ~1,000x weaker. To get usable torque back, hands can use gear reductions of 100:1 to 288:1 per joint, whereas a leg actuator typically runs around 6:1. The problem is that those ratios undermine the force transparency and simulation accuracy that manipulation depends on.

The software side is outside our usual coverage, but it’s worth naming that there are two problems in play: the hardware itself, and the intelligence that runs on it. Whatever a hand can physically do, a model still has to learn to do it, and unlike LLMs, there’s no existing body of data to train on.

Different companies are betting on different approaches, from robot teleoperation (1X) to a recently popularized approach that aims for the highest-fidelity capture of what a hand is actually doing: data collection gloves built to match a specific robot hand (companies like Sunday, Chestnut,Origami). This idea idea first grew out of a 2024 research project called Universal Manipulation Interface (UMI).

Again, we’re not experts, but the underlying idea is the same one behind Rich Sutton’s bitter lesson and Dario Amodei’s “big blob of compute”: with enough data and compute, scale wins. For humanoids, the catch is that scale depends on the hardware too - you need a hand that works, a way to collect manipulation data for it at scale, and high-fidelity data that maps back onto that specific architecture.

The Dextrous Robot Hand Landscape

Presented by Standard Propaganda: Selling Engineering Tools to Enterprises

Product-led growth sells to the user, who can’t buy anything. Founder-led sales builds on your credibility, and to a legacy enterprise, your startup has none. That’s how engineers can love your demo while the deals that would prove product-market fit to your investors never close. This walkthrough breaks down who actually signs, what they’re afraid of, and what you need to put in front of them.

The full playbook is open to readers of the newsletter—no signup required.Watch here.

More From Around the Web

Manufacturing & Startup News

  • Marathon Fusionclaims it turned mercury into stable gold using fusion neutrons at roughly 10x CERN’s 2025 rate, though it has not disclosed the quantity produced and the result remains independently unverified.

  • PreScalar raised a $7M seed round to build modular test stations that flash and verify device firmware on factory lines.

  • Firefly emerged from stealth with a $20M Series A for Ember, a platform that connects existing cameras, radios and sensors to detect incidents and coordinate a response.

  • Arda raised a $160M Series A at a $1.5B valuation for software that turns CAD files into assembly sequences, work instructions and station layouts.

Tools From Our Partners

  • Xometry - Instant quoting on custom parts, from R&D prototypes to high-volume production runs.

  • Aris Machina Protos – Design & validate anything, with the AI-native workspace for industrial science and hardware R&D.

  • SolidWorks for Startups – Full access to SolidWorks’ product development tools, free for early-stage startups.

  • Zoo – AI-native CAD with a modern geometry engine, readable code, and editable feature trees for precise, editable models.

  • Trace.Space - The AI harness for hardware R&D.

Access Every Link We’ve Ever Shared 🎞️

Refer three friends to Hardware FYI and unlock our full archive.
Missed an issue? Hunting for that one link? Every tool, guide, and article we’ve ever shared is in here.

Refer a friend

P.S. If you find something interesting—an article, tool, funding round, project, really anything—send it our way!

P.P.S. 20,000+ engineers, founders, and operators read this newsletter. If you want to reach them, get in touch.

© 2026 Benji Chia
548 Market Street PMB 72296, San Francisco, CA 94104