How Robots Learn to Feel: The Tactile Sensing Stack Explained
Tactile sensing in robots combines magnetic, force-torque, and IMU data into coordinated feedback loops that let machines handle objects and navigate spaces safely.
6 min read
0:00
0:00
Tactile sensing in robots combines magnetic, force-torque, and IMU data into coordinated feedback loops that let machines handle objects and navigate spaces safely.
Robots have long been able to see and move, but feeling, detecting contact, slip, and force at the point of interaction, has remained the hardest sensing challenge to solve at scale.
Vision tells a robot where an object is. Proprioception tells it where its own joints are. But neither tells it what is actually happening at the fingertip when it picks something up. That gap, the absence of reliable tactile feedback, is why so many manipulation tasks that look easy for humans remain unreliable for robots. According to The Robot Report, designing for a humanoid robot is considered one of the most complicated applications in robotics today, requiring coordinated management of movement, balance, vision, and reactivity across a complex web of joints, sensors, and data processing. Tactile sensing sits at the intersection of all of these. Getting it right means not just detecting contact, but detecting the right signal at the right resolution, with low enough latency to actually act on it.
Magnetic tactile sensors embed magnets in a deformable material and track field distortion to infer contact force and direction, but nearby metal structures create interference that has historically limited accuracy.
XELA Robotics uses a technology called uSkin, a magnetic tactile sensor that detects deformation in a flexible skin layer by measuring changes in embedded magnetic fields. The core physics is relatively straightforward: press on the skin, the magnets move, the field changes, and the sensor reads that change as a force vector. What is harder to solve is interference. Robots are built from metal, which disrupts magnetic fields in ways that are difficult to predict. According to The Robot Report, XELA Robotics will show at the 2026 Robotics Summit and Expo its improved magnetic interference compensation alongside uSkin integration in the Universal Manipulation Interface. That improvement matters because without reliable compensation, the sensor signal becomes noisy and unreliable the moment it is mounted on a real robotic hand, which is exactly the environment where you need it most.
The temptation is to treat magnetic interference as a noise problem you can filter out in software. The reality is more structural. The interference pattern depends on the geometry of the surrounding metal, which changes as the robot moves. A static calibration is not enough. What XELA appears to be building is a dynamic compensation model that accounts for the robot's own geometry, which is a meaningfully harder engineering challenge than standard noise filtering.
IMUs provide whole-body orientation and acceleration data while force-torque sensors capture local contact loads. Together they give a humanoid robot enough situational awareness to stay balanced while handling objects.
A single sensor type cannot cover the full sensing picture for a humanoid robot. Inertial measurement units, or IMUs, track orientation and acceleration at the body level. They tell the robot whether it is leaning, accelerating, or about to fall. Force-torque sensors at the joints and end-effectors capture what is happening at specific contact points. According to The Robot Report, this becomes critical when a humanoid robot operates in environments where it must react to unexpected contact or disturbances, not just execute pre-planned motions. The interesting design challenge is fusion: how you combine IMU data running at potentially thousands of hertz with force sensor data at different rates and resolutions, and turn that into control decisions fast enough to be physically meaningful. Analog Devices, cited in The Robot Report, is actively working on the sensor architecture side of this problem for humanoid applications.
Researchers at NTU Singapore built a magnetic seed-sized surgical robot that switches between five tools in under one second, pushing the boundaries of what force control looks like at extreme miniaturization.
Researchers at Nanyang Technological University Singapore have developed a surgical robot small enough to be compared to a seed, yet capable of switching between five different surgical tools in under one second, according to Interesting Engineering. The actuation mechanism is magnetic, which connects it directly to the broader trend of magnetic control in small-scale robots. What stands out from a systems perspective is the tool-switching speed: under one second across five configurations is a control performance benchmark that would be difficult to match with mechanical switching systems at that scale. The trade-off, as is typical with magnetically controlled micro-robots, involves external field generation, which requires equipment outside the robot body itself. The robot is not self-contained in the way a humanoid hand actuator is. But the force control principles, using field geometry to drive precise movement, share conceptual ground with what XELA is doing with uSkin at a much larger scale.
Magnetic control appears in both the NTU surgical robot and the XELA uSkin sensor, which is worth pausing on. At the micro scale, magnetic fields drive movement directly. At the humanoid scale, they encode deformation as a sensing signal. The underlying physics is the same: controlling or reading a magnetic field gradient to infer or produce mechanical action. Scale changes the engineering constraints dramatically, but the core principle travels.
Sensor density, durability, interference rejection, and latency all pull in different directions. No current system fully optimizes all four at once, and that tension is what defines the design space.
Tactile sensing for dexterous hands involves a set of competing constraints that builders have to navigate explicitly. Higher sensor density means more contact resolution, but also more data to process and more potential failure points. Flexible sensor skins conform better to complex geometries, but tend to be less durable in industrial settings. Magnetic sensing offers good resolution without requiring direct electrical contact at every sensing point, but introduces the interference problem that XELA is working to solve, as reported by The Robot Report. Latency is a constraint that cuts across all of these: the control loop for a robot hand catching or manipulating objects needs tactile feedback in single-digit milliseconds to be physically useful. Sensors that are accurate but slow are not actually useful for reactive manipulation. The Universal Manipulation Interface context for uSkin, mentioned by The Robot Report, suggests a push toward standardized integration, which would help address some of the system-level complexity if the hardware API is stable enough for third parties to build on.
Tactile sensing feeds the perception layer that sits between raw sensor data and robot decision-making. Without it, robots rely on vision and position control alone, which is not sufficient for reliable physical interaction.
The framing from The Robot Report on humanoids learning to read the room is instructive. It positions sensing not as a component-level spec question but as a system capability question. A robot that can read the room is one whose sensor fusion layer translates physical signals into actionable context fast enough to matter. Tactile data is one critical input into that layer, alongside IMU data, joint torque estimates, and vision. What is emerging from the three developments covered here, XELA's uSkin improvements, Analog Devices' humanoid sensing architecture work, and NTU's micro-robot tool switching, is a picture of tactile and force sensing maturing across multiple scales and applications simultaneously. The underlying challenge is consistent: how do you get reliable, low-latency force information from the physical world into a control loop that can act on it. The specific solutions differ by scale and application, but the engineering tension is the same everywhere.
uSkin is a magnetic tactile sensor developed by XELA Robotics that detects contact forces by measuring distortion in embedded magnetic fields within a flexible skin layer. According to The Robot Report, XELA is integrating uSkin into the Universal Manipulation Interface and demonstrating improved magnetic interference compensation at the 2026 Robotics Summit and Expo.
Robots are built from metal components that disrupt magnetic fields in geometry-dependent ways. Because the interference pattern changes as the robot moves, static calibration is not sufficient. XELA Robotics is working on dynamic compensation that accounts for the robot's own metallic structure, which is a prerequisite for reliable performance in real deployments.
IMUs measure orientation and acceleration at the body level, giving humanoid robots the data they need to maintain balance and detect unexpected disturbances. As reported by The Robot Report in collaboration with Analog Devices, IMU data must be fused with force-torque sensor data and vision to give a humanoid enough situational awareness to operate in unstructured environments.
Researchers at Nanyang Technological University Singapore built a seed-sized robot that switches between five surgical tools in under one second using magnetic actuation, according to Interesting Engineering. The tool-switching speed at that scale is a notable control benchmark and demonstrates that magnetic actuation principles are viable far below the scale of humanoid robot components.
The Universal Manipulation Interface is a standardized framework for robot manipulation hardware integration. XELA Robotics integrating uSkin into this interface, as reported by The Robot Report, signals a shift from bespoke tactile sensor deployments toward a more standardized approach, which is typically a prerequisite for broader adoption and scale in hardware markets.