Robots get a feel for touch: breakthroughs in skin, sensing, and dexterity set stage for human-like manipulation

Human touch is a marvel of biological complexity, relying on multiple specialized mechanoreceptors embedded in the skin that detect diverse stimuli such as vibration, stretch, and texture with remarkable spatial resolution. Crucially, touch is an active sense—humans constantly press, slide, and adjust to transform raw sensory input into rich perception. This dynamic, embodied interaction begins early in development, as fetuses explore their environment through touch, laying the foundation for understanding physical properties like weight and resistance, which underscores the challenge for robots to replicate not just sensation but the developmental role of tactile experience.

Replicating human-like touch in robotics demands more than just outfitting machines with sensors; it requires embedding distributed intelligence and localized processing within soft, compliant bodies that can physically interact with the environment. Inspired by biological exemplars like octopuses, which house most of their neurons in their limbs, researchers emphasize the importance of 'embodied intelligence' where the robot's body itself contributes to perception and action. This approach offloads cognitive workload from central processors, enabling robots to adjust grip and posture fluidly by leveraging the deformation and filtering properties of artificial skin and compliant materials.

By early 2026, thought leaders like Prof. Perla Maiolino at IROS 2025 highlighted a paradigm shift where the robot body is not merely a sensor platform but an active generator of data and mediator of safe human-robot interaction. Equipping robots with advanced sensors, compliant morphologies, and artificial skins transforms their bodies into perceptive surfaces that continuously gather experiential data, enabling learning through physical interaction. This embodied, sensorized approach allows robots to 'bump, grasp, yield, and recover,' mirroring the biological complexity of human touch and fostering a form of intelligence that arises through acting, sensing, and adapting in the world.

By late 2025, neuromorphic robotic e-skin (NRE-skin) emerged as a breakthrough in artificial skin technology, mimicking the human nervous system with a hierarchical architecture that enables robots to sense touch and pain and trigger instant reflexive motor responses by bypassing the CPU. This skin not only detects harmful forces but also actively monitors its own integrity by sending continuous electrical pulses when intact and halting them upon damage, signaling injury locations to the robot and its owner. Its modular design, featuring magnetic patches for rapid self-repair, enhances durability and ease of maintenance, setting a new standard for responsive and resilient robotic skin.

By early 2026, advances in soft, stretchable, and transparent multimodal sensors integrated ionic liquid optics with conductive fabrics to capture complex motions like twisting and bending, as demonstrated by Professor Sunan Co’s ultra-stretchable transparent sensors. These innovations, combined with AI-driven learning methods, enabled a single wrist-mounted sensor to decode intricate finger movements, while novel sensor architectures reduced wiring complexity by consolidating multiple sensor metrics into a single wire, facilitating scalable large-area tactile sensing crucial for wearable and robotic skin applications.

Large-area artificial skins built on flexible PCBs employing capacitive transduction have been developed to provide distributed pressure sensing across robot bodies, enabling nuanced control of interaction forces and safe human-robot collaboration. These skins confront system-level challenges such as conforming to complex 3D surfaces and managing computational bandwidth, but their integration allows robots to detect multiple simultaneous contacts, perform full-body manipulation tasks, and even navigate cluttered environments without relying on cameras by prioritizing tactile feedback to control forces and maintain robust contact during interactions.

The fusion of distributed proximity sensors—compact, low-cost time-of-flight devices—with tactile skin technologies has significantly enhanced robotic spatial awareness and safety in collaborative settings. This synergy enables robots to generate detailed point clouds of their surroundings, perform self-localization and environment mapping, and detect human presence, facilitating safe navigation and manipulation in cluttered spaces by maintaining distance from obstacles and controlling interaction forces when contact is unavoidable. As Perla Maiolino highlighted at IROS 2025, such sensorized, compliant robotic bodies transform into perceptive surfaces that generate experiential data through physical interaction, fundamentally shifting the robot’s body from a passive sensor container to an active mediator of safe, intelligent behavior.

Recent tactile sensor designs have embraced the dual-modality principle inspired by human mechanoreceptors, integrating static and dynamic capacitive sensing circuits on a single PCB layer to simultaneously capture sustained pressure and transient contact events without crosstalk, as exemplified by the TSF-85 sensor. Capacitive sensing remains favored for robotic fingertips due to its manufacturability and immunity to optical and magnetic interference, while inclusion of fingertip orientation channels provides essential spatial context to accurately interpret pressure maps when finger pose is not directly observable, collectively advancing tactile sensing fidelity critical for dexterous robotic manipulation.

By early 2026, adaptive control systems leveraging Gaussian process regression (GPR) have revolutionized robotic dexterity by enabling robots to replicate complex human grasping motions with minimal training data. This approach not only deciphers the nuanced relationship between an object's environmental stiffness and the human's force-position commands but also achieves remarkable accuracy, reducing position root-mean-square error by up to 74% in extrapolation scenarios. Such advancements allow robots to generalize manipulation skills to unfamiliar objects efficiently, marking a significant leap toward safe and precise interaction in dynamic, real-world environments.

Integrating tactile sensors across the entire robot body, including limbs and torso, has expanded the frontier of whole-body multicontact control, enabling robots to maintain balance and execute complex postures like sitting on narrow surfaces through reactive adjustments. This is achieved by combining nonlinear model predictive control with push and force control strategies, allowing robots to interpret tactile feedback and apply precise reference forces dynamically, thus enhancing stability and motion fluidity beyond traditional limb-focused approaches.

Advances presented at IROS 2025 underscore the challenge and necessity of synthesizing whole-body multi-contact motions from human data, capturing the intricate interplay of discrete contact sequences and continuous dynamics. This holistic learning approach is crucial for humanoid robots to emulate the fluidity and adaptability of human dexterity, particularly in environments demanding complex, coordinated contact interactions.

The EmArm robotic system exemplifies cutting-edge embodied sensorimotor integration by fusing a rigid skeletal framework with large-area soft tactile skins and proprioceptive sensing within a closed-loop perception–action framework. This integration achieves submillimetre tactile localization and robust whole-arm sensing, enabling real-time physical adaptation, contact-rich manipulation, and tactile-driven trajectory replanning even under visual occlusion and environmental disturbances. As detailed by Tang et al. (2026), this scalable approach paves the way for physically intelligent robots capable of safe, adaptive, and intuitive operation in unstructured, dynamic settings.

By early 2026, Cecilia Laschi's analysis underscored how embodied intelligence and morphological computation are revolutionizing soft robotics by enabling energy-frugal, robust systems that offload computational demands from centralized controllers to the robot's physical structure. This approach draws inspiration from biological intelligence, which achieves adaptive sensory-motor coordination under severe energy constraints, and envisions a future where soft robots might operate entirely without electronics, relying solely on mechanically programmed behaviors.

Research from USC revealed that decentralized control mechanisms observed in sea stars—where each tube foot independently responds to local mechanical feedback—can inspire robots capable of resilient, adaptive locomotion without a central brain. This bioinspired strategy promotes energy efficiency and robustness, enabling autonomous machines to adapt and recover through distributed decision-making, a principle that challenges traditional centralized robotic control architectures.

A collaborative effort by Korea University, ETH Zurich, and UCLA introduced a dog-training-inspired framework that allows legged robots to rapidly acquire new skills through natural human interactions such as touch, gestures, and voice commands, achieving a remarkable 97% success rate after just a few real-world trials. This method exemplifies how embodied intelligence extends beyond morphology to include interactive learning paradigms that enhance adaptability in complex environments.

Perla Maiolino’s keynote at IROS 2025 highlighted the transformative role of compliant morphologies and artificial skins in advancing embodied intelligence, where the robotic body itself becomes an active data generator rather than a passive sensor platform. This shift enables robots to continuously gather experiential data through physical interaction, fostering safer human-robot collaboration and positioning the body as the foundational mediator of intelligence and interaction.

By mid-2026, PSYONIC's Ability Hand prosthetic emerged as a groundbreaking tool for capturing high-fidelity human manipulation data, recording nuanced grip force, finger positioning, and micro-corrections during real-world tasks performed by amputees. This rich dataset, described as potentially robotics' 'ImageNet moment,' is being integrated through collaborations with NVIDIA and ABB Robotics into open learning frameworks and industrial automation platforms, enabling real-to-real transfer learning that grounds robot training in authentic human tactile and motor signals. Founder Dr. Aadeel Akhtar emphasized that using the same bionic hand for both humans and robots creates a unique data bridge, capturing subtle tactile nuances that robots currently lack and significantly advancing robotic dexterity.

Concurrently, the SnowOrigin team at Peking University, backed by prominent investors including Gong Hongjia and Lu Qi, advanced the field with next-generation EMG wristbands and panoramic headbands that capture detailed human manipulation data encompassing intention, force, and micro-control. These wearable devices overcome limitations of traditional motion capture by providing structured, high-quality data suitable for continuous long-term use, at lower cost and weight, thereby facilitating scalable embodied intelligence data infrastructure. Their pioneering real-time decoding of full hand posture from electromyography signals positions SnowOrigin at the forefront of a paradigm shift toward data-centric robot training and enhanced human-computer interaction.

By mid-2026, innovations like EleTac have demonstrated the practical potential of soft robotic grippers that mimic biological dexterity, combining gentle, versatile grasping with high-resolution tactile sensing inspired by the elephant trunk’s tip. EleTac’s integration of pneumatically actuated soft fingers and an internal fisheye camera, powered by deep learning algorithms, enables nuanced tactile feedback and adaptive pressure control, while its compact, lightweight, and cost-effective design facilitates plug-and-play deployment across industrial and service robotics, including delicate handling and assistance for vulnerable populations.

Contactile’s PapillArray tactile compute module has emerged as a leading commercial solution enhancing robotic touch by enabling real-time grip force adjustments, which significantly reduce damage risks when handling delicate or irregular objects. With adoption by over 100 global customers spanning major industrial and research organizations, Contactile’s emphasis on manufacturability and robustness positions it at the forefront of advancing safe, reliable object manipulation amid the surge in humanoid and mobile robotics investments.

The introduction of ThorArena marks a pivotal industry innovation by providing a simulation-based benchmark that rigorously evaluates humanoid robots’ ability to withstand realistic physical forces during real-world tasks—a dimension largely overlooked by conventional no-force tests. Utilizing data captured from human interactions via VR and sensor-equipped tools, ThorArena’s Force-Aware Tracking Score (FATS) metric has revealed significant performance disparities among humanoid control systems, with the Thor2 system achieving near-perfect balance across multiple tasks, underscoring the benchmark’s value in driving advancements in robot dexterity and manipulation capabilities.