Robotic hands take center stage: the race to human-level dexterity heats up

By early 2026, the robotics community reached a consensus that the primary bottleneck for practical humanoid robots was not locomotion or safety, but mastering dexterous hand manipulation. This realization marked a pivotal shift from previous decades where industrial robots thrived in controlled environments tailored to their capabilities; now, service and humanoid robots faced unstructured, cluttered human spaces filled with soft objects and awkward angles, demanding unprecedented hand adaptability. As one analysis put it, “Humanoids will not take off until we solve the hands. Not walking. Not balance. Not even safety.”

This shift in focus was exemplified by early projects such as Stanford’s ALOHA and Columbia University’s MiniBee, which moved away from scripted manipulation towards learning-based skill acquisition. Stanford’s ALOHA reframed manipulation from an intractable physics challenge to a training problem, while MiniBee quietly pursued understanding how machines can interact with the world without causing damage. These efforts underscored a broader industry transition, recognizing that the true promise of humanoid robots lies not in their humanlike appearance but in their ability to operate tools and objects designed for human hands within everyday environments.

By early 2026, the engineering of humanoid hands remains a pinnacle of mechanical complexity and cost, often surpassing the price of the entire robot itself due to intricate tendon-driven systems and advanced sensing integration. Shadow Robot, a leader in the field, exemplifies this challenge through its evolution from pneumatic muscle actuation to proprietary motor-driven techniques, striving to balance anthropomorphic fidelity with functional reliability. This high complexity also translates into formidable control difficulties, as humanoid hands with approximately 20 axes of motion demand computational capabilities comparable to managing five industrial robot arms simultaneously, often necessitating teleoperation due to current limitations in fusing force and tactile sensor data.

Practical design trade-offs increasingly favor simplifying locomotion and manipulation mechanisms to enhance usability and cost-effectiveness. Instead of replicating complex legged locomotion, experts advocate mounting humanoid upper bodies on wheeled bases and replacing multi-fingered hands with simpler grippers, while strategically investing in advanced shoulder and wrist mechanisms to maintain manipulation dexterity. Despite their critical role, shoulders and wrists remain underexplored areas in humanoid robotics, even though modular enhancements—such as adding spine degrees of freedom with tilting and pitch/yaw motions using just a few motors—offer promising avenues to improve manipulation capabilities without prohibitive cost increases.

By early 2026, the pursuit of human-like dexterity in robotic hands remains a formidable challenge, with companies such as Switzerland's Mimic Robotics and Norway's Physical Robotics leading efforts to replicate nuanced human manipulation. Mimic Robotics leverages humans wearing haptic devices to perform tasks that can be transferred to robots, underscoring the importance of teleoperation and data collection from human analogs. This approach aligns with the understanding that a high degree of freedom—such as five fingers with 10 or more degrees of freedom—is essential for complex tasks like unscrewing bottle caps or delicate tool use, which simpler robotic hands cannot reliably perform.

Embedded tactile sensing stands out as a critical enabler for advanced robotic manipulation, particularly in high-DOF hands designed for tasks like welding and assembly. However, integrating dense sensor arrays into the compact, moving joints of robotic hands presents significant packaging and wiring challenges, as engineers strive to maintain sensor integrity amid constant motion. This hardware bottleneck is compounded by the slower pace of actuator and material science advancements, with limitations such as copper resistance and rare earth magnet capabilities constraining progress despite rapid AI improvements.

Recent iterative developments highlight the ongoing evolution of humanoid robots equipped with advanced hands; for example, a Gen 2 robot recently achieved its first steps with plans to integrate its sophisticated upper body soon. This milestone reflects the broader trend of balancing hardware design and software control, where human-like hand structures not only facilitate more natural teleoperation but also improve the fidelity of neural network training by providing data that closely mirrors human manipulation.

Despite impressive strides, robots still struggle with delicate manipulation tasks that humans perform effortlessly, such as handling fragile objects without damage. This gap illustrates the complexity of translating human tactile intuition into robotic control, emphasizing that while AI models and sensor technology advance, the physical embodiment of these capabilities in hardware remains a critical hurdle to achieving truly dexterous robotic hands.

By early 2026, the robotics field recognized that innovation in robot form factors had been stifled largely due to AI limitations, but the advent of robotic foundation models promised to democratize design experimentation. Analysts suggested that if hobbyists and researchers could simply assemble a robot in their garage and load a versatile foundation model capable of controlling diverse bodies and tools, it would unleash a wave of creativity and accelerate development. This vision aligns with the idea that physical intelligence should be embodiment-agnostic, much like the human brain treats tools as extensions of the body, enabling robots to adapt fluidly to different morphologies and tasks.

Chenying Liu’s pioneering research on embodied physical intelligence, particularly through origami-inspired robots, exemplifies how embedding intelligence directly into a robot’s physical structure—via geometry, materials, and control—can dramatically enhance autonomy and robustness. Her work reveals that a robot’s form is not merely a passive shell but an active participant in sensing, processing, decision-making, and movement. By leveraging unconventional materials and foldable structures, these robots transcend traditional control paradigms, embodying intelligence that is distributed throughout their physical form rather than centralized solely in software.

By early 2026, Josie Hughes and her team at EPFL's CREATE Lab have been at the forefront of integrating embodied intelligence with AI to create physically intelligent robots that excel in robustness and adaptability. Their approach uniquely offloads computation to the robot's body, enabling machines to operate effectively across diverse environments by combining control AI with the intrinsic intelligence arising from the robot's morphology and material properties. Hughes emphasizes that this synergy unlocks capabilities unattainable by pure AI or control methods alone, marking a paradigm shift in robotic design.

Central to their innovation is the fusion of rigid and soft compliant structures inspired by biological exemplars such as the human hand, elephant trunks, and octopus tentacles. This hybrid design philosophy enables robotic manipulators to achieve both dexterity and precision through continuum structures capable of infinite bending configurations. Such soft, reconfigurable robots can dynamically morph their shape, facilitating novel locomotion and manipulation strategies that challenge traditional planning paradigms—like a compliant rover that transitions from rolling to swimming to traverse obstacles directly rather than circumventing them.

This research embraces a decentralized control model, drawing inspiration from biological systems where control is distributed—octopus arms, for instance, operate semi-autonomously from the brain. By leveraging the unpredictability and compliance of soft materials, Hughes’ work moves away from rigid kinematics toward robots that exploit environmental interactions and uncertainty to enhance adaptability. This approach has practical applications in challenging unstructured environments, including steep slope viticulture in Switzerland, underwater swimming robots, delicate agricultural harvesting, and medical devices mimicking human soft organ functions.

Augmenting these physical innovations, Hughes pioneers the use of generative AI to co-design robotic structures that fully exploit their physicality and environmental interactions. Supported by the Advanced Research + Invention Agency's Robot Dexterity programme, which aims to revolutionize robotic capabilities and boost human productivity, the CREATE Lab established in 2021 serves as a hub for advancing bio-inspired soft robotic manipulation through generative design. This fusion of AI-driven design and embodied intelligence signals a transformative leap toward robots that are not only physically intelligent but also creatively engineered for complex real-world tasks.

By early 2026, Figure Robotics had demonstrated an extraordinary pace in humanoid robot development, achieving walking capability with their first-generation Figure 01 within a year of founding—an accomplishment they believe to be among the fastest in history. This rapid progress was coupled with a deep learning curve in hand design; the initial tendon-driven hands, inspired by human anatomy with motors in the forearm, proved suboptimal. Over five generations, Figure iterated their hand designs, culminating in a high-degree-of-freedom hand that matches human dexterity, underscoring the critical role of hands in enabling robots to learn from human video and replicate complex motions—an essential step toward achieving artificial general intelligence (AGI).

The evolution from Figure 01 through Figure 03 highlights a series of thoughtful design refinements that enhanced both form and function. Notably, Figure 02 relocated the battery from a backpack to the torso, tripled computing power, and introduced an aircraft-style exoskeleton where the outer skin bears structural loads. Figure 03 further slimmed the silhouette, reduced weight, added soft foam padding, and incorporated swappable fabric clothing alongside tactile-sensing hands, collectively improving dexterity and utility. These iterative improvements reflect a maturing design philosophy focused on blending robustness with human-like adaptability.

Looking ahead, Figure Robotics anticipates a revolutionary leap with Figure 04, which they liken to an 'iPhone One moment'—a transformative shift that transcends the incremental refinements of previous models. This suggests that while Figures 01 through 03 laid a solid foundation through rapid iteration and incremental innovation, Figure 04 aims to redefine humanoid robot design in a fundamentally new way, potentially accelerating the path toward robots with human-level intelligence and physical capability.