Proprioceptive Shape Estimation of Tensegrity Manipulators Using Energy Minimisation

This paper proposes and validates a proprioceptive shape estimation method for large-scale tensegrity manipulators that utilizes inclination angles from IMUs and energy minimization to achieve high-accuracy reconstruction (2.1% error) under both static and disturbed conditions without relying on external sensors.

The Gist

Imagine you have a giant, wiggly robot arm made not of solid metal joints, but of floating sticks held together by a web of tight rubber bands. This is a tensegrity manipulator. It's like a 3D spiderweb made of rigid poles and elastic strings. Because it has no fixed hinges, it can bend, twist, and squeeze into tight spaces just like an octopus tentacle or an elephant trunk.

But here's the problem: How does the robot know what shape it's in?

In a normal robot, you can just look at the angles of its joints (like a human looking at their elbow). But in this wiggly robot, the "joints" are floating in mid-air, held up only by tension. If you can't see the robot (maybe it's in a dark cave or underwater), how does it know if it's curled up in a ball or stretched out straight?

This paper presents a clever solution to that problem. Here is the breakdown in simple terms:

1. The Problem: "Blind" Robots

Most robots use cameras or external sensors to see where they are. But cameras are expensive, heavy, and don't work well in the dark or underwater.

• The old way: Stick a camera on the robot. (Expensive, fragile).
• The new way: Give the robot a "sense of self" (proprioception), like how you know your arm is raised even with your eyes closed.

2. The Solution: The "Gravity Compass"

The researchers attached tiny sensors (called IMUs, similar to the ones in your smartphone) to every single stick (strut) in the robot.

• What they measure: They don't measure how long the rubber bands are (which is hard). Instead, they just measure which way gravity is pulling on each stick.
• The Analogy: Imagine you are blindfolded and holding a long, flexible pole. If you tilt the pole, you can feel the weight shift. If you have 20 people holding 20 different poles, and they all tell you "I'm tilting 30 degrees to the left," you can figure out the overall shape of the whole group just by listening to their tilts.

3. The Magic Trick: "Energy Minimization"

Once the robot knows the tilt of every stick, how does it figure out the 3D shape? It uses a mathematical concept called Energy Minimization.

• The Metaphor: Think of the robot's rubber bands as springs.
  • If the robot is in a weird, twisted shape, the springs are stretched or squished in unnatural ways. This creates high tension (high energy). Nature hates high tension; it wants to relax.
  • If the robot is in its "natural" shape, the springs are relaxed. This is low energy.
• The Process: The computer acts like a digital "relaxation" machine. It says: "Okay, based on the tilt sensors, let's try to arrange the sticks so that the rubber bands are as relaxed as possible."
  • It starts with a guess (maybe a random mess).
  • It calculates the "tension energy."
  • It tweaks the shape slightly to lower that energy.
  • It repeats this thousands of times until it finds the shape where the energy is the lowest.
  • Result: The shape that requires the least amount of energy to hold the sensors' tilt angles is almost certainly the actual shape of the robot.

4. The Experiment: The Giant Wiggly Arm

The team tested this on a massive robot arm called TM-40.

• It has 5 layers stacked on top of each other.
• It has 20 sticks and 80 rubber bands.
• It is over 1 meter long.

They tried three things:

1. Collapsed State: The robot was squished into a ball. The computer guessed a ball, then "relaxed" the energy until it matched the real squished shape.
2. Expanded State: The robot was stretched out. The computer guessed a stretched shape and refined it.
3. Random Guesses: They started the computer with completely random, crazy shapes. Every single time, the "energy minimization" trick pulled the guess back to the correct shape.

The Result: The computer estimated the shape with 98% accuracy (only a 2% error). It even worked when they physically pushed and bent the robot with their hands!

Why This Matters

This is a big deal because:

• It's Cheap: You don't need expensive cameras or stretchy sensors on every rubber band. Just cheap tilt sensors on the sticks.
• It's General: This method could work on any tensegrity robot, big or small, without needing to redesign the robot's body.
• It's Self-Contained: The robot can figure out its own shape anywhere, even in the dark or deep underwater, without needing to "see" anything outside.

The Bottom Line

The researchers taught a wiggly, joint-less robot to "feel" its own shape by listening to gravity and asking itself, "What shape would make my rubber bands the most relaxed?" It's a brilliant way to give a robot a sense of its own body, using nothing but a little bit of math and a few cheap sensors.

Technical Summary

Here is a detailed technical summary of the paper "Proprioceptive Shape Estimation of Tensegrity Manipulators Using Energy Minimisation."

1. Problem Statement

Tensegrity manipulators (structures composed of rigid struts and flexible cables) offer high compliance, adaptability, and load-carrying capacity, making them ideal for safe human-robot interaction. However, controlling and modeling these systems is challenging due to their infinite degrees of freedom (DOF) and lack of discrete joints.

• The Core Challenge: Accurate shape estimation (determining the 3D configuration of the robot) is fundamental for control.
• Limitations of Current Methods:
  • Exteroceptive sensors (cameras, external trackers) are costly, limited to specific environments, and prone to occlusion.
  • Proprioceptive methods often rely on measuring cable lengths. However, many tensegrity designs use passive cables whose lengths are not directly measurable, leading to incomplete data.
  • Existing Proprioceptive Solutions: While some studies use Inertial Measurement Units (IMUs), they often combine IMU data with external sensors or have only been validated on small, single-layer structures. There is a lack of proven methods for large-scale, multi-layer tensegrity manipulators that rely solely on a single sensor type (IMU) without external references.

2. Methodology

The authors propose a proprioceptive shape estimation algorithm based on energy minimization. The method reconstructs the 3D shape of the manipulator using only the inclination angles of the rigid struts relative to gravity.

A. System Model

• Hardware: The study utilizes the TM-40, a highly redundant, five-layer tensegrity manipulator consisting of 20 rigid CFRP struts, 40 nodes, and 80 tensile members (40 active pneumatic cylinders and 40 passive Kevlar cables).
• Sensing: Each strut is equipped with a 6-axis IMU (LSM6DSOX).
  • Measured Data: The inclination angle ($\phi$) of each strut relative to the gravity vector is derived from accelerometer and gyroscope data using a complementary filter.
  • Unknowns: The yaw angle ($\theta$) around the gravity vector is difficult to measure accurately due to magnetic interference in dense structures. Therefore, $\theta$ is treated as an unknown variable to be solved by the algorithm. The spatial center positions ($p$) of the struts are also unknown.

B. Mathematical Formulation

The shape reconstruction is framed as an optimization problem where the system seeks a state of minimum elastic potential energy.

1. Energy Function: The total elastic energy ($e$) of the cable network is calculated based on Hooke's law. It is the sum of the energy stored in all cables, defined by their stiffness ($K$), current length ($m$), and natural length ($b$).
   $$e = \frac{1}{2} (m_s - b_s)^T K (m_s - b_s)$$
   Note: For simplification in this work, natural lengths ($b_s$) were assumed to be zero, though the authors note this as a limitation for future work.

2. Geometric Constraints: The current cable lengths ($m_s$) are functions of the strut center positions ($p$) and orientations ($q$), which depend on the measured inclination ($\phi$) and the unknown yaw ($\theta$).

3. Optimization: The algorithm minimizes the energy function $e$ with respect to the unknown parameters $p$ (positions) and $\theta$ (yaw angles).

   • Solver: A first-order Gradient Descent (GD) method is used.
   • Update Rule: The parameters are updated iteratively:
     $$p_{k+1} = p_k - \alpha \frac{\partial e}{\partial p}, \quad \theta_{k+1} = \theta_k - \beta \frac{\partial e}{\partial \theta}$$
   • Convergence: The process continues until the gradients approach zero, indicating the structure has reached a minimal energy state that matches the physical constraints and sensor data.

3. Key Contributions

1. Scalability: The first demonstration of a proprioceptive shape estimator on a full-scale, five-layer tensegrity manipulator (1160 mm length, 20 struts), moving beyond previous single-layer validations.
2. Single-Sensor Reliance: The method achieves shape estimation using only IMU inclination data, eliminating the need for external sensors, cable length sensors, or complex strain gauges.
3. Robustness to Initial Conditions: The algorithm successfully converges to the correct shape from arbitrary initial conditions (collapsed, expanded, or random states).
4. Disturbance Handling: The system maintains stable shape estimation even when subjected to external manual perturbations.

4. Experimental Results

Experiments were conducted on the TM-40 manipulator under static and dynamic conditions.

• Static Accuracy:
  • The method estimated the manipulator's total length with an error of 2.1% (24.5 mm error on an 1160 mm structure).
  • The estimated shape converged to the actual physical shape in both "collapsed" and "expanded" starting states.
  • Convergence Speed: The algorithm reached an optimal solution in approximately 7.3 seconds (1000 optimization steps). A single update step took ~7.36 ms.
• Robustness:
  • Tested with 10 different random initial conditions. In all cases, the optimization converged to the same minimal energy state, demonstrating the algorithm's stability.
• External Disturbances:
  • When manual forces were applied to the tip of the arm, the estimator successfully tracked the resulting deformations in real-time visualization (RViz2).
• Limitations Observed:
  • Accuracy decreased during deep bending. The authors attribute this to the simplification of assuming zero natural cable lengths and the lack of a dynamic stiffness model for the pneumatic actuators.

5. Significance and Future Work

• Significance: This work proves that complex, redundant tensegrity robots can be controlled and monitored using only low-cost, easily mountable IMUs. This removes the dependency on expensive external tracking systems or complex internal cable sensing, making tensegrity robots more viable for real-world applications in confined spaces or collaborative environments.
• Future Directions:
  • Refining the Model: Incorporating the natural lengths of cables and developing a stiffness model that maps pneumatic pressure to cable stiffness to improve accuracy during deep bending.
  • Real-time Performance: Parallelizing sensor data processing to reduce latency and enable faster control loops.
  • Generalization: Applying the method to other tensegrity-based robotic systems with minimal hardware modifications.

In conclusion, the paper presents a robust, energy-based framework that solves the "open research problem" of proprioceptive shape estimation for large-scale tensegrity manipulators, paving the way for more autonomous and adaptable soft robotic systems.