Can Reversible Robotic Hands Solve Humanoid Dexterity?
EPFL engineers have built a reversible robotic hand that outperforms human dexterity in controlled tasks, while LG is already betting on dexterous hands for home robots.
6 min read
0:00
0:00
EPFL engineers have built a reversible robotic hand that outperforms human dexterity in controlled tasks, while LG is already betting on dexterous hands for home robots.
EPFL built a robotic hand that exceeds human dexterity in controlled manipulation tasks and can physically detach and move independently from its arm.
According to New Atlas, engineers at EPFL's school of engineering have developed a robotic hand they describe as capable of outperforming human dexterity in controlled manipulation tasks. That is a significant claim. The human hand has long been the benchmark that robotic engineers aspire toward, with roughly 27 degrees of freedom across bones, joints, and tendons. The fact that EPFL's team is publicly making this comparison suggests the design achieves something genuinely unusual in terms of finger coordination and grip versatility. The 'reversible' element adds another layer: the hand can reportedly crawl away from its own arm, implying a reconfigurable or detachable mechanical architecture that goes well beyond conventional end-effector design.
The term reversible here appears to refer to the hand's ability to detach from its parent arm and operate independently, crawling under its own power. From what we can see, this points to an onboard actuation and power system compact enough to fit within the hand structure itself. If accurate, this has real implications for modular robot design, where end-effectors could be swapped, repositioned, or even deployed as semi-autonomous tools.
The current state of the art in robotic hands includes systems like the Shadow Dexterous Hand, which replicates human hand kinematics closely, and simpler three-fingered grippers used in most industrial and humanoid robots today. Most commercial humanoid platforms, including early Optimus and Figure 01 hands, prioritize robust grasping over fine manipulation. EPFL's claim of surpassing human performance in controlled tasks would, if validated, represent a meaningful step beyond what is currently deployed at scale.
The reversible, crawling capability implies onboard actuation within the hand itself, likely using tendon-driven or miniaturized direct-drive motors paired with a detachable wrist interface.
From what we can see in the New Atlas report, the hand's ability to detach and crawl points to a self-contained actuation system. In high-dexterity robotic hands, actuators are typically either located in the forearm and connected via tendons or cables, or placed directly in the finger links using miniaturized motors. A hand that can operate independently almost certainly carries its own power delivery and motor control hardware onboard. The engineering challenge here is significant: miniaturizing enough actuator torque and control electronics to fit inside a hand-sized package while maintaining the degrees of freedom needed for dexterous manipulation is exactly the kind of problem that has blocked progress in this space for decades.
Remote actuation via tendons keeps weight out of the hand but introduces friction, compliance, and routing complexity. Onboard actuation adds mass to the most distal part of the arm, which increases the torque demand on wrist and elbow joints. EPFL's choice to enable full independence suggests they accepted that mass penalty in exchange for modularity and self-sufficient operation. Whether that trade-off makes sense for a humanoid platform is an open question.
LG's CLOi home robot, unveiled at CES 2026, uses dexterous hands and visual learning to automate cooking, cleaning, and chores as part of a 'Zero Labor Home' vision.
As reported by New Atlas, LG debuted its CLOi domestic humanoid robot at CES 2026. The platform is a wheeled humanoid designed specifically for household tasks including cooking, cleaning, and general chore management. LG describes this as part of a 'Zero Labor Home' vision, where advanced Physical AI handles routine domestic work. The robot uses dexterous hands combined with visual learning, meaning it interprets its environment through cameras and maps what it sees to manipulation strategies. This is a markedly different design philosophy from industrial humanoids: the home environment is unstructured, unpredictable, and built entirely around human-scale objects that require fine motor control.
CLOi uses a wheeled base rather than legs. This is a pragmatic engineering call. Wheels are mechanically simpler, more energy efficient, and far more reliable in the predictable flat-floor environments of most homes. Legs add enormous complexity and cost without delivering proportional benefit indoors. The dexterity challenge in a home robot is mostly in the hands, not the locomotion system.
According to New Atlas, CLOi uses visual learning to automate tasks. This approach, where the robot observes and imitates manipulation strategies from visual input, is central to how Physical AI companies are trying to make robots generalizable across the thousands of object types and task variations found in real homes. The hand hardware has to be capable enough to execute what the vision system requests.
Both EPFL's research hand and LG's home robot converge on the same core bottleneck: building hand actuators that are compact, powerful, and precise enough to handle unstructured real-world objects.
The data suggests these two stories are different points on the same technology curve. EPFL is pushing the frontier of what robotic hands can do mechanically, demonstrating that human-level and potentially superhuman dexterity is achievable in a lab-built system. LG is on the commercialization side, deploying the best available dexterous hand technology into a consumer product context. The gap between these two positions is the standard research-to-product journey, but in robotics that gap has historically been enormous. What is changing now is the convergence of better miniaturized actuators, improved motor controllers, and vision-based learning systems that can compensate for mechanical limitations through software.
Durability, cost, tactile sensing integration, and the gap between lab benchmarks and real-world task variability remain the primary blockers for production-ready dexterous hands.
Research hands like EPFL's are typically built for demonstration and testing rather than thousands of operational hours. Production-grade robotic hands need to withstand repeated grasping cycles, temperature variation, dust, moisture, and unexpected contact forces without degrading. Cost is another wall: high-degree-of-freedom hands with onboard actuation use many small, precision components that are expensive to manufacture at scale. Tactile sensing, knowing how hard the hand is pressing and what surface it is touching, remains underdeveloped in most systems and is critical for manipulation tasks involving fragile or irregular objects. Finally, the benchmark problem is real: performing well on controlled lab tasks does not automatically transfer to the chaos of a real kitchen or living room.
High-dexterity hands with many actuated degrees of freedom have more potential failure points than simple grippers. Each additional joint is another bearing, motor, cable, or gear that can wear or break. For home robots deployed in consumer settings, maintenance expectations are essentially zero. The hand hardware needs to be robust enough to operate for years without service intervals.
EPFL-class research hands are hand-assembled and use precision components optimized for performance, not cost. Bringing that to a price point compatible with consumer or even light industrial markets requires fundamental manufacturing changes: injection-molded structures, off-the-shelf micro-motors at volume, and simplified assembly. This is a supply chain and manufacturing engineering challenge as much as it is a design one.
Dexterous hand actuators are emerging as a critical sub-system in the humanoid supply chain, with research breakthroughs and commercial deployments arriving simultaneously in early 2026.
The convergence of EPFL's dexterity breakthrough and LG's commercial deployment signals that the hand actuator problem is moving from theoretical to urgent. Humanoid platform builders have historically focused on locomotion and gross manipulation, accepting simplified grippers as a temporary compromise. As Physical AI software improves and commercial use cases demand finer manipulation, the pressure on hand hardware is increasing. Companies like LG entering the market create demand signals that justify investment in specialized hand actuator supply chains. Research outputs from institutions like EPFL provide the technical blueprints that suppliers and startups will build toward. The two dynamics are feeding each other.
According to New Atlas, EPFL's hand can outperform human dexterity in controlled tasks and physically detach from its arm to crawl independently. That self-sufficient locomotion capability implies onboard actuation, which is unusual and suggests a genuinely novel mechanical architecture compared to most research and commercial hands.
As reported by New Atlas, LG's CLOi is a wheeled domestic humanoid designed to cook, clean, and manage household chores. It uses dexterous hands paired with visual learning and Physical AI to interpret and manipulate the objects and environments found in a typical home, as part of LG's Zero Labor Home vision.
The human hand has roughly 27 degrees of freedom and integrates touch, pressure, and proprioceptive feedback simultaneously. Replicating that in a robot requires compact actuators, tactile sensors, and real-time control systems, all packed into a small, lightweight structure that must also be durable enough for continuous operation.
The data suggests we are in a transition period. LG is already deploying dexterous hands in a consumer product announced at CES 2026, while EPFL is demonstrating what the upper performance boundary looks like. The gap between lab capability and mass-market cost and durability likely spans three to seven years for high-dexterity systems.
From what we can see, the crawling and self-contained behavior points to onboard miniaturized motors, likely either direct-drive or tendon-driven with small gearboxes. The specific motor topology is not confirmed in the source reporting, but the performance claims suggest high torque density relative to the actuator size.