Clone Robotics posted a three-part thread shortly after 18:26 UTC on 1 October 2026: first a 41-second clip of a teleoperated Torso 3 hand, then a note that the hand was built in April and is the last of the old generation, and finally a Torso 4 launch set for November. What it showed, in other words, is hardware the company is retiring.
Key takeaways
- The thread went up on 1 October 2026 between 18:26 and 18:32 UTC
- The hand teleoperation video runs 40.8 seconds
- The Torso 3 hand was built in April 2026 and is "the last of Clone's legacy Hand designs"
- Over six months the company redesigned the entire system from the ground up, optimising for "simulatibility and controllability"
- Torso 4 is billed as a human-level torso meant to automate stationary bimanual tasks for enterprises
The demo shows the past, not the future
The footage documents teleoperation, not autonomy — a human is driving the hand. The hardware on screen is also six months old and, by the maker's own account, closes out a design line. So the clip says nothing about what Torso 4 can actually do. Forty-one seconds is also far too short to judge grasp repeatability.
The value of the thread lies elsewhere: in the stated reason for the redesign. Clone does not claim the new build is stronger or more dexterous. It claims the new build is easier to simulate and control.
Two schools of hand design, one goal
On the same day Boston Dynamics unveiled a new hand for its Atlas robot — four fingers, 13 degrees of freedom, deliberately no pinky, built around reliability and manufacturability. Clone goes the other way: human anatomy remains the template.
| Criterion | Clone Torso 3 → 4 | Atlas GR3 |
|---|---|---|
| Design template | human hand anatomy | deliberately non-anthropomorphic |
| Finger count | full human layout | four, no pinky |
| Stated design goal | ease of simulation and control | reliability and manufacturability |
Despite opposite philosophies, both teams now reach for the same closing argument: the hand has to be predictable for the software learning to use it. That is a shift, because a year ago dexterous manipulation was still sold on degree-of-freedom counts.
Designing for the training pipeline
"Simulatibility" in practice means a claim that the hand can be faithfully reproduced in a simulator — and therefore that a policy trained in simulation stands a chance on real hardware. That is the sim-to-real problem, solved here on the mechanical side rather than in the algorithm. A simpler, more deterministic build means a narrower gap between simulation and reality.
Why it matters
The yardstick for a robotic hand is changing. It is no longer resemblance to a human hand but how cheaply the thing can be taught a task. Designers are starting to build for the software's requirements rather than for anatomy. Clone is doing that while keeping the human template, which is the harder route — and which will show whether anthropomorphism actually buys anything once training cost is what counts.
What's next?
- Torso 4 is set for November 2026 — only then can anyone judge what the robot does unaided versus under supervision
- Clone has not published degree-of-freedom or force figures for the new hand, so any comparison with Atlas stays qualitative
- The "stationary bimanual task" scope marks a narrow niche: no locomotion means deployment at a workstation, not across a plant
