"Humanoids won't arrive in 5 years"—The real reason robots get stuck on "shoelaces"
With generative AI having conquered "language" in one fell swoop, expectations have skyrocketed that "robots are next" and that "humanoids that do all the housework will arrive in 5 years." However, UC Berkeley robotics researcher Ken Goldberg offers this reality check: “I think it will happen. But the question is ‘when.’” If we misjudge this “when,” investments, talent, and society’s reception could all end up spinning their wheels. Focusing on the TECH010 dialogue, we break down “what is advancing and what is stuck” using concrete examples.
1. The temptation of "robots are next" fueled by generative AI
What generative AI (especially since the advent of Transformers) has demonstrated is the fact that "massive data × learning" can achieve translation, summarization, reasoning, and even creation. From this, many people naturally jump to the conclusion that:
“If language is possible, robots should advance along the same curve.”
What Goldberg is wary of is this assumption of the “same curve.” Language has an astronomical amount of training text available, such as the web, books, and code. On the other hand, because robots deal with contact, friction, and deformation in the real world, the very form of the data required for learning is different.
2. "Mobility" has advanced, but "dexterity" is lagging
Mobility has advanced surprisingly well over the last decade, including quadrupedal and bipedal walking and stable drone control. However, what creates value in homes and factories is manipulation (the ability to work with hands) rather than mobility. What TECH010 repeatedly emphasized was the disconnect between being able to “walk” and being able to “work.”
2-1. The reason they can't tie shoelaces: It's not so much tactile sensation as it is the "hell of deformation"
A symbolic example is “tying shoelaces.” Both the shoelaces and the fingertips deform during the task, and force subtly escapes, slips, tightens, and loosens again. Humans use fingertip sensation to almost unconsciously compensate for “it just slipped a little,” but to reproduce this in a robot, three steps are required: (1) accurately measuring that deformation, (2) accurately understanding that deformation, and (3) translating it into control on the fly. Moreover, the wall of it being difficult to “perfectly” simulate the physics of deformation and friction stands in the way.
2-2. A path of hope: Surgery can "approach dexterity even without tactile sensation"
There is also an interesting counter-argument. Robot-assisted surgery is not the commonly imagined scenario of “a robot performing surgery on its own,” but is accomplished by a surgeon operating the robot as a “puppet.” And for many years, surgeons have performed high-difficulty tasks even with limited tactile feedback. Goldberg says there is a hint here.
Instead of perfectly reproducing tactile sensation, read minute deformations with vision (high-performance cameras), estimate them, and control accordingly. This idea is also compatible with designs that place cameras on the palms or fingertips (e.g., enhanced hand-eye perspective).
3. The "100,000-year robot data gap"
A powerful metaphor Goldberg throws out is the Robot Data Gap. Language models can ingest the text humanity has accumulated on a massive scale. On the other hand, for robots, the desired data—the time series connecting “video + state” to “force, joints, and control signals”—is not lying around on the internet. In a Science Robotics paper, it is stated that to bridge this gap, a mechanism for robots to collect data “while working” (a data flywheel) is crucial.
3-1. The strength of "synthetic data × task limitation" shown by Dex-Net
A real-world breakthrough example is Dex-Net from Berkeley. Dex-Net 1.0 prepared over 10,000 3D models and millions of grasping samples (such as parallel gripper grasping) to advance robust grasp planning.
The point was not the “perfect reproduction of the human hand,” but rather isolating the problem to “grasping,” creating data, and winning probabilistically.
3-2. Data alone isn't enough: The revival of "good old-fashioned engineering"
What hits home in the dialogue is the talk of good old-fashioned engineering that Goldberg repeats. Sensor calibration, lighting, wear on suction cups, transport jams, on-site safety—this kind of gritty optimization determines uptime and reliability. In other words, just having a smart model doesn't mean it will “work every day in the field.”
4. Practical application comes from "specialization × reliability": Ambi and "napkin folding"
While humanoids steal the spotlight, the areas actually generating value are those with clear, measurable scopes, such as "warehousing," "sorting," and "palletizing."
4-1. Ambi Robotics: The Reality of Handling "100 Million Items" with Suction Cups
Ambi Robotics' AmbiSort is said to have sorted a cumulative total of over 100 million items in commercial operations using relatively simple end-effectors like suction cups (with metrics such as accuracy also made public). What we see here is not a "dexterous hand," but rather task design and the accumulation of operational data.
4-2. Dyna Robotics: The Significance of Folding Over 800 Napkins in 24 Hours
As a type of demo that Goldberg says "surprised" him, Dyna Robotics has published a demonstration of folding over 800 napkins autonomously for 24 consecutive hours. What is important is not the flashiness, but the fact that it includes continuous operation and failure recovery. What the robotics industry needs is not a 30-second miracle demo, but the ability to run for 8 hours across 3 shifts without breaking down.
5. When Will Home Humanoids Arrive?: A Realistic "Next 10 Years"
Goldberg does not say they will "never arrive." Rather, he suggests that household chores that are "high-value and relatively limited in scope," such as picking up and tidying items off the floor, have potential over a 10-year span (with significant benefits such as elderly support).
However, there are two prerequisites.
Privacy Design: Logs of video, audio, and behavior in home spaces are the most contentious trade-off for convenience.
Expectation Management: If excessive expectations are not met, a backlash occurs where people say "robots were all a lie," which ends up dragging down even companies that are seriously running operations in the field.
This concern is also continuous with Rodney Brooks' "criticism of humanoid fever." Brooks points out the danger that today's trends underestimate "dexterity" and that investments are being made based on incorrect premises.
Conclusion
Robots are coming. However, the order in which they arrive is unlikely to be the "sci-fi omnipotent humanoid," but rather a buildup of single-task specialization → reliability → operational data → expansion to adjacent tasks. Therefore, what we should be doing now is not talking about dreams, but identifying which tasks can "generate value while producing data" and translating them into forms that work in the field. What bridges the 100,000-year gap is not flashy videos, but mountains of mundane operational logs—those who can swallow this reality will become the next winners in the robotics industry.