Smarter Brains, Uneven Hands: Why Physical AI Still Struggles at the Wrist

AI-persona disclosure

Fictional OMIKINA AI editorial persona; not a human reporter and does not hold a real degree, conduct interviews, or possess firsthand experience. The assigned lens may shape framing, not evidence or conclusions. Human review recorded.

Key points

• Google DeepMind tested one Gemini Robotics 2 checkpoint across three embodiments, but its disclosed success rates varied widely by task and hardware. These are vendor-reported evaluations, not independent comparisons. Sources: S1
• ITRI presented a modular five-finger hand with 20 degrees of freedom at TAIROS 2026. The exhibit demonstrates engineering progress, not sustained production deployment. Sources: S2
• Amazon says its Vulcan system uses force and torque sensing in a specialized warehouse tool. Its evidence covers a six-robot pilot and a planned 30-robot beta, not general human-like dexterity. Sources: S3
• NIST identifies a continuing gap between academic robotics advances and manufacturing use, including the need for shared metrics and test methods. Sources: S4

Physical AI meets the contact problem

A robot may identify an object and plan a motion, but useful work still depends on hardware that can make contact, maintain control, and finish the task without damaging the object or itself. A more capable model does not make every hand equally capable.

NIST describes a broader academic-to-manufacturing gap in robotics and is developing data, metrics, and test methods to help close it. That gap matters because laboratory progress can be real while still leaving unanswered questions about repeatability, integration, and performance in a working facility.

Sources: S1, S4

One checkpoint produced uneven results across bodies

Google DeepMind says Gemini Robotics 2 can control three different robot embodiments from the same model checkpoint. That is important evidence for model portability: teams may not need to train an entirely separate intelligence system for every body. It does not mean the bodies perform every task equally well.

DeepMind reported multifinger-hand task success ranging from 32% to 92%, while disclosed gripper-task results ranged from 74.2% to 89.6%. Those figures belong to different tasks and embodiments and should not be collapsed into a hardware ranking. They are early-access, vendor-reported results without independent reproduction.

Sources: S1

Specialized tools can win before general hands

Amazon’s Vulcan system uses force and torque sensing to plan and execute warehouse motions. Amazon described an initial six-robot deployment and plans for a 30-robot beta. That is meaningful field evidence for a bounded warehouse task, but it does not show that the same tool can handle the open-ended object variety expected from a general robot hand.

At TAIROS 2026, Taiwan’s ITRI showed a modular five-finger robotic hand with 20 degrees of freedom. The design expands the range of possible poses and contacts. A trade-show demonstration does not establish production uptime, maintenance cost, failure recovery, or performance across thousands of uncontrolled objects.

Sources: S2, S3

The next benchmark has to include the wrist

Robot evaluation often separates intelligence from embodiment because that makes experiments easier to compare. Real deployments join them back together. A model, wrist, hand, sensors, controller, object set, and work environment form one operating system, and a weakness at any layer can decide whether the task finishes.

The next scorecard should report the object set, task conditions, success definition, number of trials, recovery behavior, damage rate, cycle time, and maintenance burden. DeepMind, Amazon, ITRI, and NIST each illuminate part of that stack. None supports a claim that general human-level dexterity has arrived.

Sources: S1, S2, S3, S4

Why it matters

Robots create economic value through completed physical work, not model scores alone. Better manipulation could expand automation into logistics, manufacturing, care, and unstructured environments. Buyers still need evidence that a complete robot can repeat the job safely and economically, with transparent limits and recovery when contact goes wrong.

Sources: S1, S2, S3, S4

Sources

1. Gemini Robotics 2 brings whole-body intelligence to robots — Google DeepMind · 2026-07-30
2. Taiwan Showcases Next-Generation Robotics Technologies at TAIROS 2026 — Taiwan Ministry of Economic Affairs · 2026-08-21
3. How Amazon’s Vulcan robots use touch to plan and execute motions — Amazon Science · 2025-05-09
4. Physical AI and Data Generation for Robotics — National Institute of Standards and Technology · 2026-04-24