Embedded Proprioception for Autonomous Soft Robots

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 218 papers22 conceptsfrozen 2026-10-04

Traditional robots are rigid, making their movements predictable and their joint positions easy to measure with simple encoders. Soft robots, however, trade this predictability for compliance and safety. Because their bodies can bend, twist, and wrinkle in infinite ways, knowing where a soft robot is in space—or whether it has touched something—usually requires a room full of expensive external motion-capture cameras.

To move soft robots out of the lab and into the real world, researchers are developing "proprioception": an internal sense of self. By embedding sensors directly into the robot's "flesh" or using the physics of the body itself as a computational element, these systems can perceive their own shape and environment without external help.

The Body as a Computational Sensor

The most elegant way to give a robot proprioception is to treat its physical structure as a sensor. In soft systems, every deformation is a response to either an internal command or an external force. If you have a good enough mathematical model of the robot's elasticity, you don't actually need tactile sensors to "feel" the world.

Zach J. Patterson and Daniela Rus demonstrated this in Modeling and Control of Intrinsically Elasticity Coupled Soft-Rigid Robots. By leveraging the "intrinsic elastic coupling" within a robot's structure, they showed that external forces acting on unactuated parts of the robot can be estimated using only internal joint angle measurements. Essentially, the robot’s body acts as a transducer, converting environmental interactions into measurable joint displacements.

A similar logic applies to fluid-driven robots. Fan Zhu and Yilin Wen found that monitoring the internal air pressure of pneumatic actuators allows a robot to detect contact and even object slippage. In Failure Handling of Robotic Pick and Place Tasks With Multimodal Cues Under Partial Object Occlusion, they used a momentum observer to filter this pressure data, allowing a gripper to "feel" a successful grasp even when its own fingers blocked its camera's view.

Sensorized Skins and Deep Learning

While physics models are powerful, the high-dimensional "squishiness" of soft materials often defies simple equations. To capture complex 3D configurations, researchers like Ryan L. Truby have turned to sensorized "skins."

In Distributed Proprioception of 3D Configuration in Soft, Sensorized Robots via Deep Learning, Truby's team wrapped soft robotic arms in kirigami-inspired conductive films. As the arm bends, these films stretch, changing their electrical resistance. Because the relationship between resistance and 3D shape is non-linear and complex, they used Long Short-Term Memory (LSTM) networks to translate these signals into a real-time map of the robot's posture. This approach effectively replaces a $100,000 motion-capture suite with a few dollars' worth of laser-cut plastic and a neural network.

Turning Actuation Noise into Feedback

One of the biggest challenges in embedded sensing is "cross-talk." When a soft actuator moves, it inevitably stretches the sensors embedded within it, creating "noise" that can be mistaken for an external touch.

Recent work by Sergio Decherchi on the SoftMag system flips this problem on its head. Instead of trying to eliminate the interference between the actuator and the sensor, the system uses a neural network to decouple the two. By understanding how the actuator's own movement affects the magnetic tactile sensors, the robot can subtract its own "self-movement" to isolate external forces. This allowed a soft gripper to evaluate the firmness of delicate apricots with high precision, despite the actuator itself constantly shifting the sensor's baseline.

Achieving True Autonomy

The ultimate goal of this research is to cut the "visual umbilical cord" that ties robots to laboratory infrastructure. For unconventional designs like tensegrity robots—which move by buckling and rolling—this is a matter of survival.

Tzu-Yuan Lin and colleagues developed a state estimator for these robots that relies entirely on internal IMUs and motor encoders. As described in Tensegrity Robot Proprioceptive State Estimation with Geometric Constraints, they used an Invariant Extended Kalman Filter to fuse these internal signals, allowing the robot to track its position in dark or occluded environments where cameras would fail.

Whether through Fabian Schwab’s flexible capacitive sensors for swimming fish or Gregory Xie’s integrated tactile feedback for in-hand manipulation, the field is moving toward a unified goal: robots that don't need to be watched to know where they are. By embedding the "sense of self" into the material fabric of the machine, soft robots are finally becoming capable of autonomous operation in the messy, uninstrumented real world.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 16 papers
  2. Version 1 · 2026-10-04 · 18 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_proprioceptive_sensing_gap_6293b781,
  title = {proprioceptive sensing gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proprioceptive-sensing-gap/rdr_6293b78122c94ff4891145f03c46d693},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_embedded_proprioception_for_autonomous_s_6293b781_v1,
  title = {Embedded Proprioception for Autonomous Soft Robots},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/embedded-proprioception-for-autonomous-soft-robots/rdr_6293b78122c94ff4891145f03c46d693/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}