Robot Hands vs Legs: Why Hands Are the Hard Part

Robot hands vs legs is not a close contest, and autumn 2026 proved it. On September 15, IEEE Spectrum reported on Agility Robotics’ Digit 5, a 129-kilogram humanoid designed to work more than 20 hours a day in industry. Two weeks later, on October 1, the same publication reported that Boston Dynamics had redesigned the Atlas hand from scratch, with its director of robot behavior saying the team was still “figuring out what needs to change in the fine details of our design.”

The legs are a product. The hands are a research project. That gap has six concrete causes, and none of them is about computing power.

Key Takeaways

• Walking has one objective: keep the centre of mass over the support and do not fall. Grasping has no single objective, because the right grip depends on the object, the next step and what the person wanted.
• Walking has a handful of repeatable contact modes. Manipulation has a combinatorial number of them, and every change of contact is a discontinuity in the physics.
• Hands pack more independent joints into a smaller space. The new Atlas hand has 13 degrees of freedom and the Shadow Dexterous Hand has 20 actuated degrees of freedom across 24 joints.
• Grasping depends heavily on touch, which is the hardest sense to build. Boston Dynamics worked around this by using back-drivable joint motors as force sensors instead of fingertip sensors.
• Moravec’s paradox named this in 1988: tasks requiring conscious reasoning are easier for computers than unconscious sensorimotor skills.

Robot Hands vs Legs: Six Reasons the Gap Exists

One: the objective function. Walking has a single, measurable goal. Keep the centre of mass over the support polygon, keep moving forward, do not fall. You can write that as a number and optimise it. Grasping has no equivalent number. The correct way to pick up a mug depends on whether you are about to drink from it, hand it to someone, or put it in a dishwasher. The objective is defined by a future intention, which the robot does not have access to.

Two: contact modes. This is the deepest technical reason and the one most often skipped. When you walk, your foot is either on the ground or off it. Add heel strike and toe off and you have maybe three or four distinct contact states, repeating in a cycle. When you manipulate an object, the contact state is which parts of which fingers are touching which parts of the object, and whether each contact is sticking, sliding or rolling. The number of combinations explodes. Worse, each transition between contact modes is a discontinuity: the equations governing the motion change the instant a finger lifts. Control theory handles smooth systems well. It handles systems that switch equations hundreds of times per second much less well.

Three: joint density. A biped’s balance problem is dominated by a small number of large joints working together, which is why the inverted pendulum model gets you remarkably far. A hand crams many small independent joints into a space the size of a sandwich. The new Atlas hand has 13 degrees of freedom. The Shadow Dexterous Hand has 20 actuated degrees of freedom across 24 joints, driven by 20 motors or 20 antagonistic pairs of air muscles in the forearm. Every degree of freedom is another thing to actuate, sense, control and repair.

Four: sensing. Legs get excellent proprioception, which is relatively easy to build: an encoder on every joint tells you where the limb is. Hands need tactile sensing, and skin is hard. Pressure sensors on fingertips are fragile, expensive and wear out. This is exactly why Boston Dynamics’ solution was to use back-drivable motors in the joints as their own force sensors, which the IEEE Spectrum reporting describes as reacting to force and contact. It is a clever workaround for a sense nobody has built well.

Five: the variability of the world. The floor is approximately a plane. It is occasionally sloped, occasionally slippery, occasionally has a step. That is a small space of possibilities and a robot can be tested across most of it. Objects are not a small space. Thin, flat, soft, slippery, granular, deformable, fragile, hot, sticky, already-held-by-someone. A hand that works on a box tells you nothing about whether it works on a plastic bag.

Six: training cost. Legs can be trained and tuned against a physics simulator that models the ground well. Contact-rich manipulation is where simulators are weakest, because friction, deformation and impact are the hardest things to simulate accurately. OpenAI’s 2018 study on learning dexterous in-hand manipulation used reinforcement learning on a Shadow Dexterous Hand with randomised physical properties, including friction coefficients and object appearance, precisely because the simulation could not be trusted to match reality. That technique exists because the underlying problem is hard.

Now the analogy, with the mechanism in place. Walking is like riding a bicycle: one continuous skill, one failure mode, and once your body has it you have it forever. Grasping is like cooking: thousands of small situation-specific judgements, no single success criterion, and being good with a knife tells you little about being good with dough.

The Comparison in One Table

Property	Legs and walking	Hands and grasping
Objective	One, measurable: do not fall	Depends on the object and on intent
Contact modes	Three or four, cyclic	Combinatorial, with discontinuities
Independent joints involved	Few large ones	Many small ones, 13 to 20 in current designs
Key sense	Proprioception, easy to instrument	Touch, hard to build and fragile
World variability	A mostly flat plane	Effectively unbounded
Simulation quality	Good	Weakest exactly where contact matters
2026 status	Shipping industrial product	Being redesigned from scratch

The last row is the honest summary. A 129-kilogram biped is being sold to run more than 20 hours a day, while a leading robotics company is still deciding how many fingers a hand should have.

Why This Surprises Everybody

Because it is backwards from human experience, and that has a name.

Hans Moravec wrote in 1988 that “it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.” Steven Pinker summarised it in 1994: “the main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard.”

Moravec’s explanation was evolutionary. “Encoded in the large, highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it,” he wrote, while abstract thought “is a new trick, perhaps less than 100 thousand years old.” Skills that feel effortless feel that way because enormous amounts of dedicated machinery run them below awareness. The difficulty of reverse-engineering a skill tracks how long evolution spent on it.

Which gives a neat prediction, and it holds. Bipedal walking is a comparatively recent adaptation in the hominin line. Grasping hands are far older across primates. Moravec’s rule predicts that the older skill is the harder one to rebuild, and that is what engineers find.

How to Teach Your Kid About This

Ages 5–8: Walk the line, then pick up the rice

Materials: masking tape, a spoon, a small pile of rice or dried lentils, two cups.

First, tape a line on the floor, two metres long, and have your child walk it heel to toe. Then backwards. Then with their eyes nearly closed. It is easy, it is repeatable, and they will get bored.

Second, pour twenty grains of rice onto a plate and have them move the grains one at a time into a cup, using only a spoon held in a closed fist. Time it. Then try with two fingers. Then with chopsticks or two pencils.

Say the comparison out loud: walking got boring in two minutes, and the rice is still annoying. “That is exactly the problem robots have.”

Ages 9–12: Count the contact modes

Materials: one mug, one sheet of paper, a pencil.

Ask your child to list every distinct way a hand can hold the mug. By the handle. By the rim with two fingers. Around the body with a whole hand. Upside down by the base. Two hands. Pinched at the rim with a thumb and one finger. Most kids reach eight to twelve and keep going.

Then ask them to list every distinct way a foot touches the ground while walking. Heel down. Whole foot. Toe pushing off. Foot in the air. Four.

Write the two numbers next to each other. Twelve versus four is already a gap, and then explain that a robot has to decide between the twelve and predict what happens in each, while the four repeat in the same order forever. That ratio is the difficulty ratio, derived by a ten-year-old from a mug.

Ages 13+: Measure how much of grasping is touch

Materials: a tray of ten varied objects, a blindfold, thick winter gloves, a notebook.

Round one, normal: pick up each object, record success and time. This is the baseline, and it will be near-perfect.

Round two, blindfolded: same ten objects, placed in the same positions. Record success and time. Most people do surprisingly well, because touch carries them.

Round three, gloves on and eyes open: same objects. This is the interesting one. Thick gloves remove fine tactile feedback while leaving vision intact, and performance usually drops more than people expect, especially on thin, flat and soft objects.

Three numbers per object, thirty data points. The conclusion writes itself: vision tells you where things are, touch tells you what is happening. Robots have excellent vision and poor touch, which is precisely the configuration round three simulated.

The question to ask: “Why can a robot walk up stairs but not fold a shirt?”

What to Do at Home

Give credit where the difficulty actually is

When a video shows a robot doing a backflip, the honest reaction is that locomotion has got good. When a video shows a robot folding a towel, that is the harder achievement even though it looks less impressive. Teaching a kid to invert the impressiveness ranking is a small act of technical literacy that will keep paying off.

Use the contact-mode idea on other things

It explains a lot. Why catching is harder than throwing. Why writing with a pen is harder than running. Why surgery is hard and weightlifting is not, in the robot sense. Any task where the number of ways things can touch is large is a task machines find difficult, and that one heuristic predicts a surprising amount about which jobs automate first.

Watch what ships, not what demonstrates

Digit 5 is being sold for more than 20 hours of daily operation. The Atlas hand is being redesigned with an eye to making 100,000 units a year. Those two sentences tell you where each capability sits on the maturity curve, and reading product announcements for that signal is more informative than reading them for capability claims.

Let your kid be bad at a fine-motor task on purpose

Chopsticks, knitting, soldering, origami, shuffling cards. The frustration is the lesson, and it is a lesson that makes the robotics news comprehensible rather than magical. A child who has spent forty minutes failing to pick up a single grain of rice with chopsticks has a visceral understanding of why Boston Dynamics employs a director of robot behavior.

What not to do

Do not tell your kid that robots will never fold laundry. That is not what the evidence says. The evidence says manipulation is a harder problem than locomotion and is receiving enormous engineering attention right now, with real progress in both learning methods and hardware. Overconfident predictions in either direction age badly. “Harder, and being worked on” is both accurate and more interesting.

What to Watch For Over the Next 3 Months

• Week 4: Look for any manipulation result reported with a success rate across a varied object set, rather than a single successful attempt. Success rates on varied objects are the honest metric, and their scarcity in coverage is itself informative.
• Month 2 red flags: A demonstration video with no indication of how many attempts it took. Claims about dexterity that mention degrees of freedom but not reliability. Both are signs that the capability is not yet a product.
• Month 3 self-check: Ask your kid which is harder for a robot, climbing a ladder or buttoning a shirt, and why. If the answer mentions how many ways the fingers can touch the button, the contact-mode idea has landed. If it mentions strength or balance, run the mug activity again.

Frequently Asked Questions

Why is walking easier for a robot than picking things up?

Walking has one measurable objective and a handful of repeating contact states. Grasping has no single objective, because the right grip depends on intent, and a combinatorial number of contact states, each transition changing the governing physics. The difficulty is structural, not a matter of processing power.

Did robots solve walking?

Not entirely, but far enough to sell. Agility Robotics’ Digit 5 is designed for more than 20 hours of daily industrial operation, and IEEE Spectrum’s September 15, 2026 coverage focused on safety rather than on whether it can walk. When the conversation moves from capability to safety certification, a technology has matured.

What is Moravec’s paradox in one sentence?

It is the observation, stated by Hans Moravec in 1988, that tasks requiring conscious reasoning are comparatively easy for computers while unconscious sensorimotor skills are extremely hard, because the latter rest on far more evolutionary engineering.

Is touch really the bottleneck?

It is one of several, and it is the one with the fewest good solutions. Fingertip pressure sensors are fragile and wear out, which is why Boston Dynamics instead used back-drivable joint motors that react to force and contact. That is a workaround rather than a solution to artificial skin.

Which household task will robots manage last?

Nobody knows, and anybody who claims to is guessing. The structural prediction from contact modes is that the hardest tasks involve soft, thin or deformable objects with no fixed shape: folding fitted sheets, handling plastic bags, separating wet laundry. Boxes and rigid objects are much further along.

Should my kid study mechanical or software engineering for this?

The interesting work sits between them. The Atlas hand story is a mechanical design decision made for manufacturing reasons that changes what the software has to do, and the OpenAI manipulation work is a software approach built to compensate for hardware limits. Programmes that keep both, including mechatronics and robotics engineering, map onto the real problem better than either alone.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. IEEE Spectrum. (2026, October 1). “Atlas Robot’s New Hand May Outperform Humanlike Designs.” https://spectrum.ieee.org/robust-robot-hand
2. Ackerman, E. (2026, September 15). “Digit 5 May Be the First Humanoid Robot Worker That’s Truly Safe.” IEEE Spectrum. https://spectrum.ieee.org/humanoid-robot-safety
3. Wikipedia contributors. “Moravec’s paradox.” Wikipedia. https://en.wikipedia.org/wiki/Moravec%27s_paradox
4. OpenAI, Andrychowicz, M., Baker, B., Chociej, M., et al. (2018). “Learning Dexterous In-Hand Manipulation.” https://arxiv.org/abs/1808.00177
5. Wikipedia contributors. “Shadow Hand.” Wikipedia. https://en.wikipedia.org/wiki/Shadow_Hand
6. Wikipedia contributors. “Humanoid robot.” Wikipedia. https://en.wikipedia.org/wiki/Humanoid_robot
7. Occupational Safety and Health Administration. “Robotics.” US Department of Labor. https://www.osha.gov/robotics

Related reading on HiWave Makers: why the Atlas hand beats the human design, why humanoid robots fall over, and how robot dexterity gets measured.