Is it possible to have a robot achieve human-level dexterity just by watching video tutorials? To answer these questions, I talked to Rodney Brooks, the co-founder of iRobot and one of the leading researchers in the field of robotics. His take? It’s “pure fantasy thinking” to think that robots can become humanoid helpers when billions of dollars have been invested in training them to be so. Currently, robots have ‘coordination issues’ to say the least, and are ‘totally unqualified’ to simulate human touch.
The human hand offers an engineering dream of 17,000 low‑threshold mechanoreceptors that respond to minute pressure and vibrations, packing stronger on the tips of the fingers. They project to 15 different families of neurons, allowing for subtlety and adaptation of command. To achieve this in robotics requires touch sensors of breathtaking resolution although, as Brooks has noted, “we do not have such a tradition for touch data.” Though there has been success on big speech and vision corpora in AI, touch corpora remain small, making it difficult for robotics to learn about manipulation in the same way humans learn.
Musk’s Tesla, together with the AI-robotics startup Figure, is also placing a strong wager on teaching humanoids how to perform tasks by watching videos of human performance. Tesla’s Optimus project has actually transitioned from its use of motion capture suits to a vision-only system, with helmet cameras and backpack cameras capturing images of workers folding t-shirts or picking up objects. Konstantinos Laskaris, a hardware director at Tesla, went to the extent of saying, “It seems unbelievable, but our robot is learning new tasks directly from human videos!” However, Brooks points out that if robots lack force feedback, fine finger control, or rich tactile sensing, this method of training is merely simulating surface movements, not the actual underlying physics that gives human skill its flexibility.
The divide is easily observable when comparing the current humanoid hands with the most advanced tactile systems such as the F-TAC Hand, with the ability to accommodate 17 vision‑based tactile sensors across 70 percent of its surface at a spatial resolution of 0.1mm. Notably, the biomimetic approach provides closed-loop control functionality that responds in real-time during the grasp process by adjusting its strategies in order to prevent collisions and maintain stability. This is beyond the capabilities achievable with video training for humanoids.
Brooks is also skeptical of the form factor of humanoid design itself. He tries to predict that in 15 years, successful robots will “look nothing like humans,” probably including wheels, multiple arms, and perhaps five-fingered hands, but designed for mobility and manipulation, not for their human-like looks. New approaches in mobility technology, such as dual-degree-of-freedom reconfigurable wheels, already show that a non-humanoid design can be more effective than a biped design at dealing with actual environments by climbing stairs and walking through complex terrains efficiently and safely.
On the other front, engineering tactile sensors is moving towards flexible and multimodal sensors that can integrate different mechanisms such as piezoresistive, capacitive, piezoelectric, and triboelectric into one sensor for high sensitivity, large dynamic range, and robustness to environmental noise. Such sensors mounted on robotic hands will potentially replicate the human sense of touch in a multimodal manner and help with more complex manipulation, thereby emphasizing the argument that dexterity requires high-bandwidth sensory-motor cycles, not mere vision.
What is so attractive about humanoids such as Optimus and Figure 01 is their potential for plug-and-play replacement of human labor in already established processes. Yet as Brooks highlights, without resolving the problem of tactile intelligence, it is likely that robots shall only amount to staged demonstrations. In reality, as an engineer knows, manipulation in unstructured space requires sensing and control frameworks inconceivable under today’s big-tech approaches. What every tech-savvy analyst knows is that it is only via advances in tactile intelligence and adaptive mobility that better robotic capabilities are achievable.
