On October 1, 2026, Clone Robotics released video of a teleoperated Torso 3 robot hand and said it plans to unveil the fully redesigned "Torso 4" in November. The hand in the video was built in April using the previous design. It is not a demonstration of the new Torso 4.
For its next-generation machine, the company says it prioritized simulation accuracy and ease of control. The question is how a design that uses artificial muscles and tendons to reproduce humanlike motion can be turned into automation of two-armed work for businesses. To judge its practicality, it helps to separate three things: the performance of the artificial muscle on its own, the performance of the hand as a whole, and the ability to complete an actual work process.
An April-built hand is shown, with Torso 4 due in November
The company describes the video as "teleoperation of the Torso 3 hand." It is a demonstration of a human operating the machine, not evidence that the robot assessed its surroundings and carried out a task on its own.
According to a follow-up official post, the hand was built in April 2026 and is the last machine to use the previous hand design. Over the following six months, the company says, it redesigned the entire system from scratch, aiming to make simulation and control easier while maintaining humanlike capability.
However, whether the redesign maintains capability equal to or better than before will need to be confirmed through future demonstrations and test results.
The Torso 4 teaser, presented as the result of that redesign, describes a torso-type robot aimed at automating "tasks that use both arms from a fixed position" for businesses. The company named November as the timing; the post gives no price or customer shipment date.
In short, the Torso 3 video and the Torso 4 teaser should be read separately: one shows how the previous machine moves, and the other signals what the next-generation machine is meant to do.
Artificial muscle performance is not the same as hand capability
What drives Clone Robotics' hand is its proprietary artificial muscle, "Myofiber." The official hand page describes a structure in which muscles and tendons are arranged in opposition and move in response to external forces. Handling tools made for humans is one design goal, and the index finger was given enough force to pull a drill trigger, the company says.
One reason for adopting a structure close to the human hand is to make existing tools easier to use. Gripping a tool handle while operating the trigger with another finger requires more than simply pinching an object; several fingers must play different roles.
Clone Robotics aims for a hand that can handle things made for humans as they are, including the shape of the thumb and palm. Still, the general specifications on the hand page cannot be treated as the performance of the unreleased Torso 4.
The Android overview also explains a structure in which artificial muscles and tendons are manufactured as one piece and attached at positions close to those in a human skeleton. It lists the conditions needed to reproduce properties close to mammalian skeletal muscle, and the company claims Myofiber meets them.
The figures break down as follows.
| Condition in the official description | What the figure describes | What the figure alone cannot tell you |
|---|---|---|
| Response time under 50 ms | Response of the artificial muscle alone | Total latency including communication and control from operator to robot hand |
| Contraction of 30% or more with no load | Contraction rate of an unloaded artificial muscle | Range of motion and positional accuracy while gripping an object |
| Contraction force of 1 kg or more from 3 g of muscle fiber | An indicator of the force generated by the artificial muscle alone | The grip strength of the whole hand or the weight the robot can lift |
The figures are based on the company's general description as checked on October 3, 2026. "kg" is as written in the original and is not a number indicating the mass of an object the robot can lift. Nor are the figures results measured by an independent third party.
The response time under 50 ms, no-load contraction of 30% or more, and contraction force of 1 kg or more from 3 g of fiber that Clone Robotics cites mainly describe the characteristics of the artificial muscle alone. They are not figures for the latency seen when teleoperating Torso 3, nor for the probability that Torso 4 will succeed at two-armed work.
Even if the artificial muscle itself contracts quickly and strongly, whether fingertip position and force can be controlled as intended while holding a tool is a separate matter. Two-armed work also requires coordination, such as one hand supporting a part while the other operates a tool with the right timing and force.
When companies decide whether to adopt such a robot, how reliably it completes a work process matters more than the performance of the artificial muscle on its own.
The difficulty of controlling artificial muscles as intended
Artificial muscles driven by fluid pressure require modeling that includes material deformation and friction. In the typical McKibben-type artificial muscle, an outer braided structure converts the force of a pressurized inner tube expanding into contraction force along its length.
The published abstract of a modeling paper by Bertrand Tondu in 2012 identifies the nonlinear deformation of the inner material and friction in the braided section as key challenges.
Friction in the braid is also considered one of the main causes of hysteresis, where the relationship between input and output changes depending on the immediately preceding movement. This means that a simple proportional relationship, in which a fixed increase in pressure produces a proportional increase in contraction, is hard to rely on for control.
This research concerns McKibben-type artificial muscles in general. It does not directly measure the internal structure or control error of the current Myofiber. Clone Robotics has also not disclosed which parts or structures it changed in this redesign.
Even so, it is easy to see why a robot combining many flexible artificial muscles would set ease of control as a design goal. It is not enough to make a muscle contract by a target amount; that movement must show up as the intended fingertip position and contact force.
The same challenge applies to simulation. To transfer motions learned in a virtual environment to a real machine, the simulated hand and the real hand must behave sufficiently alike in response to external forces and contact with objects.
Clone Robotics' announcement indicates it is trying to improve that correspondence. However, it has not disclosed what simulation or learning methods it uses, or how much performance was obtained when motions were transferred to the real machine.
Clone and Atlas: robot hands with different design philosophies
In an official explanation of Atlas's new hand, Boston Dynamics explains why it adopted a four-finger, 13-degree-of-freedom configuration.
By driving each joint directly and omitting the pinky, the design eliminates three additional actuators and the associated cost, space, and sources of failure.
While Clone Robotics pursues a structure close to the human hand using artificial muscles and tendons, Atlas takes the direction of simplifying the structure while retaining the dexterity needed to handle tools.
Still, the number of fingers or closeness to the human hand's shape alone cannot determine which is more dexterous or practical. This is not a performance comparison under identical conditions, but a difference in design philosophy as read from each company's official explanations.
Boston Dynamics, for its part, also emphasizes simulation that accurately reproduces real-machine behavior and reinforcement learning that transfers learned motions to the real machine.
Even though the hand structures differ, reproducing motions learned in simulation reliably on a real machine is a shared technical challenge. The "ease of simulation and control" that Clone Robotics cites for Torso 4 should likewise be evaluated from this standpoint of reproducibility.
For two-armed work, the amount of human intervention and repeatability matter
The "two-armed work from a fixed position" that Torso 4 targets is an application that sets aside problems such as walking and moving to a work site, and focuses on the work ability of arms and hands.
An example would be holding a part with one hand while using a tool with the other. This is only an illustration of Torso 4's intended use, not a task Clone Robotics has announced it actually succeeded at.
To evaluate the company's goal of "automating any task that uses both arms from a fixed position," specific tasks must first be shown.
Which part is picked up, from where, moved to what position, and at what point is the task judged complete? Once such conditions are clear, success rates and task times can be measured and compared with humans and existing industrial robots.
How much human intervention is needed also matters. Whether a person only gives instructions at the start, must correct the robot's posture after failures, or teleoperates the whole process makes a large difference in the staffing required after deployment.
The Torso 3 hand shown this time is moved by human teleoperation. The video therefore confirms that it can move its fingers in fine detail, but it does not show that it can complete a work process without human operation.
What to watch for in Torso 4, planned for November, is which specific two-armed tasks it can perform, at what success rate, how many times in a row it can repeat them, and how much human intervention is needed along the way.
If continuous operating time, the durability of the artificial muscles and tendons, and the burden of replacement and maintenance are also shown, it will be easier to judge how close Clone Robotics' humanlike design has come to being a robot that businesses can use continuously on the job.