Traditional rigid robots track their state using precise joint encoders. Soft robots, however, possess continuous, theoretically infinite degrees of freedom. To track their own shape or detect environmental contact, researchers historically relied on external motion-capture systems or cameras. As noted by Xie et al. and Berdica et al., this reliance creates a sensing gap that confines soft robots to highly controlled laboratory environments.
To achieve true autonomy, a soft robot must possess intrinsic proprioception—the ability to sense its own shape and external interactions entirely from within. By treating the flexible body itself as a computational transducer, recent research demonstrates how internal strain, pressure, and structural compliance can be mapped directly to 3D kinematics and external forces.
Reconstructing Geometry from Distributed Strain
How can a continuous, squishy material know its own 3D configuration? The core intuition is to embed flexible, low-cost sensors directly into the robot's structure to capture local material deformations, then use these sparse measurements to infer global shape.
Truby developed a framework that wraps soft robot arms in a "skin" of laser-cut, kirigami-inspired sensors. By pairing these sensors with a Long Short-Term Memory (LSTM) network, the system maps localized strain directly to a 3D kinematic configuration without requiring line-of-sight cameras. Similarly, Schwab et al. embedded flexible capacitive sensors into soft actuators to provide real-time deformation feedback for closed-loop control.
[Local Material Deformation] ---> [Embedded Strain/Capacitive Sensors] ---> [Deep Learning / Analytical Model] ---> [Global 3D Kinematics]
To scale this to high-fidelity 3D meshes, Francis et al. demonstrated that deep learning models trained in physics simulators can map internal strain sensor data directly to vertex positions, allowing accurate shape estimation even during complex, contact-rich manipulation tasks. Alternatively, analytical structural models can bridge this gap: Hess combined embedded strain gauges with Euler-Bernoulli beam theory to reconstruct the real-time swimming gait and body shape of a soft robotic fish.
Because soft materials have infinite degrees of freedom, placing sensors everywhere is impractical. To solve this, Spielberg et al. introduced a neural architecture that co-learns the task control policy alongside the optimal, sparse placement of strain gauges, ensuring maximum state awareness with minimal hardware.
Fluid Cavities as Barometric Proprioceptors
For pneumatically or hydraulically actuated soft robots, the internal fluid channels themselves can double as sensors. When a soft actuator bends or compresses, the internal void volume changes, which directly alters the internal fluid pressure.
Vicari et al. leveraged this mechanism by embedding pressure sensors at the base of soft underwater tentacles. An LSTM network learns to interpret these internal pressure fluctuations to reconstruct global deformations like tip deflection and traveling wave indices.
Beyond shape tracking, this pressure-angle relationship can be mathematically modeled to turn actuators into diagnostic tools. For example, Shi et al. modeled the pneumatic pressure-angle dynamics of a 3D-printed soft rehabilitation glove. By analyzing how input pressure maps to bending angles, the authors accurately estimated patient finger joint stiffness without requiring external mechatronic rigs.
The Body as a Force Transducer
One of the most elegant aspects of soft robot proprioception is that shape estimation and force sensing are deeply coupled. Because soft materials are inherently compliant, any external force causes a physical deformation. Therefore, if a robot can accurately model its unperturbed state, any deviation in its internal sensors can be treated as a direct map of environmental interactions.
Truby formalized this via data-driven disturbance observers, treating the robot's physical body as a force sensor that infers external contact without dedicated tactile hardware. In pneumatic systems, Zhu et al. applied a generalized momentum observer to the actuator's physical model, using internal air pressure monitoring to estimate external contact forces and detect object slippage during occluded pick-and-place tasks.
External Force Applied ---> Physical Body Deforms ---> Internal State Deviation ---> Force & Slip Estimation
This concept extends to hybrid and structural systems as well:
- Elastic Coupling: Patterson et al. demonstrated that internal elastic coupling within soft-rigid hybrid structures allows external forces on unactuated segments to be calculated solely from joint angle measurements.
- Inverse Dynamics: Cangan et al. utilized a Finite Element Method (FEM) inverse dynamics solver to estimate external disturbances on a fiber-reinforced manipulator using only internal orientation sensors.
- Compliance Proxies: By leveraging compliance, as shown by Yoo et al., visual or internal deformation data can be translated directly into actionable force feedback for delicate tasks like hair care.
Closing the Loop in Unstructured Environments
Ultimately, internalizing these sensing modalities allows soft and unconventional robots to achieve robust state estimation where vision fails. For instance, tensegrity robots—which move by rolling and deforming—traditionally require external tracking.
The authors of Lin et al. overcame this by treating the contact points of a rolling tensegrity structure as temporary stationary points. By combining a contact-aided Invariant Extended Kalman Filter (InEKF) with internal geometry reconstruction, they achieved precise inertial navigation without cameras. By turning the physical compliance and internal fluid dynamics of soft bodies into computational assets, this paradigm bridges the proprioceptive gap, moving soft robotics from lab-bound novelties to truly autonomous agents.
Key Foundational Work
Data-Driven Disturbance Observers for Estimating External Forces on Soft Robots (2020) — contributed 2 concept elements to this direction.
Failure Handling of Robotic Pick and Place Tasks With Multimodal Cues Under Partial Object Occlusion (2021) — contributed 2 concept elements to this direction.
Distributed Proprioception of 3D Configuration in Soft, Sensorized Robots via Deep Learning (2020) — contributed 1 concept elements to this direction.