
Tech • AI • Robotics • Game
Three humanoid robotics efforts are highlighting diverging paths to commercialization: Clone Robotics is redesigning its muscle-driven Torso platform, Robo Party is promoting an open-source biped built for resilience, and Agibot says it has placed retail service robots in 100 stores in China.
Clone Robotics released new footage of Torso 3, a humanoid upper-body-and-leg system suspended from an overhead rig and moved through wide, coordinated motions under human teleoperation. The robot’s shoulders, arms, torso, hips, legs, and hands move together in a way that appears unusually lifelike because the machine is built around a human-style skeleton, artificial muscles, tendons, and fluid control rather than conventional joint motors and gearboxes.
Most humanoids from companies such as Tesla, Figure, and Unitree rely on electric motors placed at the joints. Clone instead uses what it calls a musculoskeletal design in which contracting artificial muscles pull on bones through tendons, producing motion that more closely resembles biological movement. That approach may improve natural motion, but it also makes modeling and control significantly harder.
Clone said it has spent the last six months redesigning the system from the ground up and now considers parts of the Torso 3 hardware part of a legacy design. The next platform, Torso 4, is scheduled to launch in November with a clearer commercial target: automating stationary bimanual enterprise tasks, meaning work done in one place with two hands at a workstation.
The redesign centers on making the robot easier to simulate and easier to control while preserving what Clone describes as human-level capabilities. That shift reflects a wider robotics trend: skills are increasingly trained first in simulation before being transferred to real hardware. For a robot driven by muscles, tendons, and pressurized fluid, accurate simulation is especially difficult, making this redesign strategically important.
A close-up teleoperation demo posted on October 1 showed Torso 3’s five-finger hand, built with artificial muscles and tendons arranged to mimic human anatomy. Clone said the hand dates to April 2026 and is the last of its legacy hand designs. One important upgrade is joint-angle feedback, which gives the controller direct information on where finger joints actually moved after muscle commands, a key requirement for precise grasping.
On the same day, Boston Dynamics revealed new hands for Atlas, underscoring two different engineering philosophies. Boston’s design uses four fingers, 13 degrees of freedom, and a motor in every joint, omitting a pinky to reduce cost, size, and failure points. Clone is pursuing a full five-finger anatomy to maximize natural human-like grip range, even at the cost of greater complexity.
Despite the striking motion demos, Clone has not published performance metrics, task benchmarks, or evidence that Torso 4 can autonomously complete workplace jobs. The latest demonstrations are teleoperated, not autonomous, and the company has not detailed how much human intervention the platform still requires. November is expected to clarify whether the human-body-inspired design can translate into durable industrial performance.
At IROS 2026 in Pittsburgh, Robo Party introduced RP1, also called Roboto01, with a public “Kick Me” demo that invited attendees to push or kick the robot to test its balance recovery. The system uses Party OS and a training framework called UFO, which applies unsupervised reinforcement learning so the robot can develop motor skills through trial and error rather than pre-recorded motion playback.
RP1 builds on the company’s earlier open-source humanoid project RPO, launched in January 2026 and credited with more than 2,500 GitHub stars. Robo Party says it has a global community of more than 5,000 developers and plans to gradually release motion control systems, simulation environments, SDKs, and training tools. The company is positioning RP1 as a full-stack open platform rather than a closed proprietary robot.
Agibot said on September 28 that it had deployed 100 X2 humanoid robots across 100 stores operated by cookware brand ASD in China. The robots, dressed in store uniforms, greet shoppers, ask about household size, cooking habits, and budget, recommend products, and guide customers around the store. In one example, the robot suggested a wok for stir-frying under 1,000 yuan, about $140, then handed the customer to a human employee when promotions were discussed.
Humanoid robotics is moving beyond spectacle toward specific commercial tests, but the evidence remains uneven. The next milestone will be whether these systems can prove autonomy, reliability, and business value outside carefully managed demonstrations.
Ask a question