Soft sensors are the "nervous system" of modern robotics, providing the compliance and sensitivity needed for delicate tasks like handling eggs or performing surgery. However, unlike rigid sensors, soft materials have a "memory." When you compress a silicone skin and release it, the material does not snap back instantly; it crawls back.
This phenomenon, known as hysteresis, creates a fundamental disconnect between physical reality and sensor output. If a robot's finger is still slowly expanding from a previous grasp, a sensor might report a "ghost" pressure that isn't actually there. To build truly responsive soft robots, researchers are moving away from static snapshots of data and toward models that account for the material's history.
The Physical Burden of Memory
In rigid electronics, a specific input (like pressure) usually maps to a specific output (like voltage). In soft systems, the output depends on whether the material is currently being loaded or unloaded.
K.D. Wise identified this in early humidity sensors, noting that sensing films often retain characteristics of their previous state, leading to measurement errors during rapid fluctuations. In tactile sensing, C. C. Geschke observed that mechanical fatigue and hysteresis cause a "shifting baseline," where the sensor fails to return to zero after use.
This behavior stems from viscoelasticity—a combination of fluid-like (viscous) and solid-like (elastic) properties. As Wei Hong explains, this nature decouples instantaneous deformation from applied force. The material exhibits "creep" (continuing to deform under constant load) and "stress relaxation" (decreasing internal stress over time), making a simple one-to-one mapping of signal-to-force impossible.
Hardware Solutions: Structural Bypasses
One way to solve hysteresis is to prevent it at the material level. K.D. Wise achieved a 1.0-second response time by replacing flat films with thousands of micron-scale polyimide columns, reducing the bulk material's ability to "trap" its state. Similarly, Fuhua Xue and colleagues developed a "gradient stiffness sliding" strategy that allows sensors to remain highly linear and virtually hysteresis-free by changing how the internal structures interact during stretching.
Other researchers, like Perla Maiolino, have swapped traditional silicones for industrial 3D fabrics. These materials are more resistant to the "ghost signals" caused by temperature and mechanical lag, providing a more stable foundation for the iCub humanoid's skin.
Software Solutions: Modeling the Timeline
While hardware can minimize hysteresis, it rarely eliminates it. The more dominant research direction involves temporal compensation: teaching the robot's "brain" to understand the material's "memory."
1. Dynamic Baseline Adaptation
If a sensor's baseline shifts because the material is slow to recover, the software must track that shift in real-time. Aiden Swann and Matthew Strong developed the TensorTouch framework, which uses dynamic baseline adaptation. By continuously updating the "reference image" of a vision-based tactile sensor, the system accounts for the gel's failure to return to its exact original state, maintaining sub-millimeter precision over long periods of use.
2. Integrating Historical State
Rather than looking at a single frame of data, newer models treat sensor readings as a sequence. Siqi Shang and Yuke Zhu demonstrated this with the FORTE gripper. To compensate for the stress relaxation of flexible polymers, their model integrates historical pressure data. This allows the gripper to handle fragile objects like raspberries by "knowing" how much of the current pressure signal is real force and how much is just material lag.
3. Velocity and Acceleration Awareness
Wei Hong found that incorporating the rate of deformation—velocity and acceleration—into force estimation models significantly reduces errors. By understanding how fast a material is moving, a neural network can predict the viscoelastic lag and "subtract" it from the final reading.
Bridging the Sim-to-Real Gap
The complexity of these time-dependent behaviors makes them notoriously difficult to simulate. Seung Hyun Kim and Jiawang Guo note that standard physics-based simulations often fail to capture non-linear damping, leading to a gap between how a robot performs in a virtual world versus the real one.
To fix this, researchers like Carmelo Sferrazza and Raffaello D'Andrea are using high-fidelity Finite Element Method (FEM) simulations to generate synthetic data that includes these dynamic properties. By training neural networks on simulations that "expect" hysteresis, the resulting models are much more robust when deployed on physical hardware.
Why It Works
The shift from static to temporal modeling works because it acknowledges a physical truth: softness is inseparable from time. By treating a sensor reading not as a point in space, but as a point in a trajectory, researchers like Zhuo Chen and Lorenzo Jamone are creating "sensor-invariant" models. These models don't just see the current deformation; they understand the physical process of the material, allowing robots to "feel" with the same temporal nuance as a human hand.