Why robot hands are still hard — and how independent eval makes progress measurable

Every few months a dexterous hand demo floods timelines: cards shuffled, eggs cradled, bottles uncapped. The hardware is real. The progress is often real. The measurement standard usually is not.

Most deployed robots still use parallel-jaw grippers for a reason. Hands are expensive, fragile, sensing-poor relative to human skin, and algorithmically unforgiving. Without independent evaluation, the field oscillates between awe and amnesia.

The hardware reality (without the mythology)

A human hand is not just “more fingers.” It is roughly two dozen degrees of freedom coupled with dense tactile sensing, compliant tissue, and decades of embodied priors. Research-grade dexterous hands can cost as much as a car, still lack human-comparable touch density, and introduce failure modes grippers never see: tendon stretch, fingertip slip under micro-vibration, and contact plans that look perfect until an object’s mass distribution disagrees.

That is why the industrial default remains the gripper. It is also why hand demos deserve scrutiny rather than cynicism: the problem is hard enough that highlight reels are inevitable — and insufficient.

Demo reels optimize for shareability, not science

A shareable demo selects for:

• favorable objects,
• favorable lighting and camera angles,
• operator-in-the-loop resets,
• and retries that never appear on disk.

A scientific grasp evaluation selects for:

• frozen protocols,
• declared object sets and mass properties,
• failure taxonomies,
• and statistics that survive a skeptic.

If your team only ships the first, you cannot tell whether next quarter’s “better hand” is better hardware, better teleop, better editing, or better luck.

A failure taxonomy beats a single success rate

Dexterous manipulation fails in distinct ways. Lump them together and you will “fix” the wrong subsystem.

Useful categories include:

• Approach / pre-grasp errors — the hand never reaches a stable grasp geometry.
• Slip after lift — grasp forms, then shear or rotation defeats friction.
• Overload / force saturation — actuators hit limits before the task completes.
• Contact-planning errors — model predicts a contact the real surface does not support.
• Regrasp thrash — recovery behaviors consume the trial without task progress.
• Object-property mismatch — mass, COM, or friction outside the training support.

Report counts per mode. A policy that fails by slip wants different remedies than one that fails by planning.

Where simulation helps — and where it lies

Simulation is the only affordable way to sweep object mass, friction, and pose distributions at scale. It is also an excellent liar if you let it grade itself.

Haga’s policy stress pillar uses adversarial mass/friction tiers on robosuite tasks such as Lift, Stack, PickPlaceCan, and Door, reporting success with Wilson confidence intervals and showing severe-tier failures. That is not a dexterous-hand leaderboard. It is the evaluation shape hands need: perturb contact physics, keep seeds honest, and refuse gate-only PASS language. See lift stress and methodology.

For hands specifically, extend the same discipline:

1. Define grasp-success separately from post-grasp transport success.
2. Sweep mass and friction even when mesh assets look identical.
3. Log contact events and fingertip force proxies when available.
4. Keep a held-out object set that marketing does not get to rehearse.
5. When hardware arrives, reuse the same taxonomy — do not invent a friendlier one.

Sim evidence vs hardware evidence (keep them labeled)

Claim you want	Evidence that can support it	Evidence that cannot
Policy survives contact noise in sim	Adversarial tier sweeps with CIs	One teleop demo
Hand design improves stable lift rate	Hardware trials under frozen protocol	Different objects each week
System is pilot-ready in a cell	Hardware-in-the-loop on that cell’s SKUs	Sim success alone

The Stanford AI Index (2026) gap — about 89.4% sim manipulation success versus about 12% on real household tasks — should hang on the wall of every dexterous-manipulation standup. Shared vocabulary without shared measurement is how fields stagnate while looking busy.

What “independent eval” adds for hands

Independence is not moral theater. It is a practical antidote to test overfitting:

• thresholds you cannot quietly loosen,
• failure galleries you cannot quietly delete,
• and methodology a partner can read without joining your Slack.

Haga’s wedge is sim-first and private: physics verification for policies and for world-model priors that increasingly sit behind manipulation stacks. We do not claim to replace academic hand benchmarks, and we do not operate a public leaderboard. We make claim-bounded packets teams can share.

If generative video or learned worlds are in your grasp pipeline, world-model physics QA belongs in the same conversation — see Physics-IQ lab evidence. A beautiful hand policy trained on inconsistent contact video is still learning fiction.

A measurement checklist for the next demo cycle

Before you cut the next public video, answer:

1. What is the exact task definition and timeout?
2. Which objects are in-distribution vs held-out?
3. What mass/friction assumptions were declared?
4. What is the failure taxonomy and raw counts?
5. How many trials, and what interval accompanies the rate?
6. Was teleop or autonomous control used — and is that labeled?
7. Which claims are sim-only, which are hardware, which are neither yet?

If you cannot answer those in writing, you are not ready to say “solved.”

Next step

Hands will keep being hard. That is fine. Unmeasured hardness is not.

Read Haga’s methodology, browse the lab, or request an eval if you have a manipulation policy or world-model scene that needs an independent stress test.