The Unhandy Valley

The name is inspired by The Uncanny Valley, and the curve follows a similar trajectory. There are many qualities that go into the axes of complexity (intricacy, controllability, part count, cost, maintainability) and usefulness (fine manipulation, tactile feedback, tool use, repeatability, strength). I’m focusing this comparison on an end effector’s ability to generalize to diverse tasks and environments. If you only want to do a few tasks, then just choose the simplest gripper for the job.

The Gripper Maxima

So what actually matters for an end effector? A key quality is opposition — the ability to meet at a point to create antagonistic forces on an object. This is very simple for a gripper, since its linear motion drives the two halves together, meaning it can pick up both small and large objects.

With just enough complexity, you can end up with a reliable, repeatable, and scalable solution. Opposition is inherent to the approach. Add a rigid “nail” and you can pick and pry away. Soften the inner jaws and gain strong conformal grips to various geometries. It just makes sense.

Grippers with one degree of freedom (DoF), like the SO-101 and YAM, are great starting points. With some refinement, you end up with designs like Generalist and Dyna. Throw in more DoF, and you end up with Sunday’s Memo or Reward AI’s gripper. Add much more than that, and you can actually start losing usefulness.

The majority of tasks can be done with these simple approaches.

The Valley

This is for those who shoot for the moon (the human hand), but struggle to take off. Following biomimicry but lacking complexity is the worst design choice. The issue really comes back to the original point of opposition, since the kinematics and shape of the fingers have to be carefully designed so that they can meet at various points. Without enough DoF, a hand cannot form good pinches. Yes, it can pick up larger objects, but so can grippers…

Unitree’s Dex3-1 occupies an awkward middle ground between a gripper and a hand. The three-fingered approach without splay makes thumb-finger pinches difficult. The linkage based Pollen Robotics Amazing Hand is cheap and simple, but its kinematics also lack good fingertip opposition. Same goes for the Psyonic Ability Hand. Switch to bulky in-finger motors, and you get the LEAP hand, which has impressive optical tactile sensors, but its size and kinematics make it unwieldy. Pivot to a tendon drive like TheRobotStudio’s Dexhand (my first hand :D), and you end up with a design that’s almost impossible to build and suffers from both durability issues and poor kinematics.

One of the common arguments for biomimetic hands is the ability to use human tools — a gripper can’t pick up and use a drill. It’s similar to how the latest frontier models have advanced their computer use abilities for navigating interfaces made for humans, even though it’s much more efficient to build programmatic hooks for direct control. But the hardware world moves much slower than this. By the time humanoid hands become both advanced enough and commoditized, it makes more sense to have integrated tools specific for robots.

If you do not have the resources to escape the Valley, I’d recommend you turn around and climb back up to The Gripper Maxima.

The Crawl to Biomimicry

Ok, now you have reached a design that does most of it right:

• Fingers can oppose each other at various locations
• A soft outer shell, ideally covered with tactile sensors
• A conformal palm
• Tendon-driven or direct-drive actuation

Following the kinematic chain of a human hand does not mean your robot moves like one. Depending on who you ask, our hands have 20–27 DoF. Others believe it is senseless to put a number on it because every joint is compliant in all directions, and the algorithms behind evolution added these degrees for a reason. I think of this as input DoF (how many knobs can you turn) and DoM (degrees of motion) out. One input can drive multiple coupled or compliant motions.

On the Crawl, you hit a decision point. Do you pursue biomimicry all the way, or strategically deviate?

Let's go through those who choose to pursue. First, we have Figure, whose hand is shown to be quite durable, but doesn't have enough mobility for dexterous tasks. I have yet to see a Figure robot do something that a gripper cannot do.

Now, let's look at the in-finger motor hands, Wuji and Sharpa. These are a great example of where accuracy and repeatability matter more than strength for dexterous tasks. Because both are available as platforms to build upon, most of the leading research uses them. There are many more demonstrations of what they can do compared to the walled gardens of the big humanoid players.

Next up, the tendon trifecta: 1X, Tesla, and Foundation. Tendons are incredible as they make the fingers easily backdrivable and powerful (at the cost of a huge complexity jump). Additionally, moving the actuation away from the fingers puts physics in your favor, as it's much easier to dissipate heat from a larger surface area. Tendons also let you freely mold the shape of the fingers and palm without the constraint of cramming your actuators inside.

As I was writing this essay, the new Boston Dynamics Atlas GR3 hand came out. This is the only design in this section that strategically deviated from the human hand. It strikes a great balance between complexity and dexterity. Having only 4 fingers simplifies the design without giving much up (sorry pinkie). The larger finger size and kinematics of the thumb means that working with small objects will be harder, which is where more biomimetic hands like Wuji and Sharpa win out.

So What’s Next?

It’s incredible how adaptable the human hand is. Everyone is born with base stats, and based on how you use your hand, you “level up your build” and can end up being able to scale mountains, repair electronics with microscopic components, or anything in between. We essentially “RL” both our hardware and models as we grow with our hands.

My advice to those building end effectors — use humans as inspiration, but don’t copy blindly. See what you can simplify without losing out on usefulness. Also, don't assume that our hands are perfect; why not add an extra thumb? If you are still driven to pursue biomimicry, then accept that simplifying joints with rigid motion destroys the compliance that makes the human hand so capable. There are research efforts (like Daxo and Clone) to create hands with compliance from first principles, but with current actuation technology, it’s not scalable yet.

The machines of tomorrow will use end effectors with many different form factors. Most will have simple grippers that are just good enough. Other cool approaches, like suction based, hydraulic driven, or soft robotics, will find their own niches in the market. Most dexterous hands will be direct drive and converge on taking just enough biological inspiration. These designs do not need to be as good as humans to accomplish the majority of tasks. For a few, this won’t be enough. Novel actuation technology that uses tendons to drive a compliant hand will be the only way to reach human-level performance. Many will be stronger, have more DoF, or use novel sensing, but none will be as good as our hands across all categories. To actually create a human hand, you cannot translate biology into mechanics; you have to replicate it. Then at that point, ask yourself, is it even worth it??

Regardless of what hardware you think is best, today’s embodied models still drive hands like they are grippers. Scaling language and video models alone will never reach the full potential of an embodied hand. Instead, we need models that understand the symmetries of our world, the equal and opposite reactions.