💻 computer science

From Impact to Insight: Dynamics-Aware Proprioceptive Terrain Sensing on Granular Media

This paper presents a dynamics-aware proprioceptive sensing framework that overcomes the limitations of quasi-static assumptions by explicitly modeling acceleration-dependent added-mass effects, enabling robots to accurately estimate granular terrain properties during high-speed hopping.

Original authors: Yifeng Zhang, Yue Wu, Jake Futterman, Jacob Meseha, Eduardo Rosales, Irie Cooper, J. Diego Caporale, Feifei Qian

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yifeng Zhang, Yue Wu, Jake Futterman, Jacob Meseha, Eduardo Rosales, Irie Cooper, J. Diego Caporale, Feifei Qian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to walk across a beach. If you walk slowly and carefully, your feet sink in just a little bit, and you can easily tell how soft the sand is. You might think, "Okay, the sand is soft here, so I need to lift my foot a bit higher next time." This is how most robots currently try to "feel" the ground: they assume that if they move slowly, the ground behaves predictably, like a sponge.

But what happens when you start running or hopping across that same beach?

When you hit the sand at high speed, the physics change completely. The sand doesn't just act like a soft sponge; it acts like a fluid that gets pushed aside, or even like a heavy weight that suddenly attaches itself to your shoe. If your robot is still using the "slow walking" rules to guess the ground conditions while it's hopping fast, it will get the math wrong. It might think the ground is harder than it is, or softer, leading to a stumble or a crash.

This paper is about teaching a robot to understand the difference between "slow walking" and "fast hopping" so it can feel the ground accurately, even when it's moving at high speeds.

The Problem: The "Slow Walker" Mistake

The researchers built a small, one-legged robot (a hopper) and made it jump on a bed of glass beads (which act like sand). They found that when the robot jumped fast:

The "Slow" Rules Failed: The robot's internal sensors (which measure motor current and position) thought the ground was behaving one way, but a real force sensor on the foot showed something totally different.
The Speed Trap: The faster the robot jumped, the bigger the mistake became.
The Hidden Culprit: It wasn't just the speed that confused the robot. It was the acceleration. When the robot's foot hit the ground and had to stop suddenly, or when the robot's controller suddenly stiffened the leg to push off, the sand didn't just resist; it added extra "weight" to the foot.

The Analogy: The "Heavy Backpack" Effect

Think of it this way:

Quasi-Static (Slow) Model: Imagine walking through a crowd. If you walk slowly, you just push people aside gently. The force you feel is proportional to how many people you push.
Dynamic (Fast) Model: Now imagine sprinting through that same crowd. You aren't just pushing people aside; you are dragging a whole group of them with you for a split second. They act like a heavy backpack that suddenly appears on your shoulders the moment you start running.

The researchers discovered that when a robot foot hits sand fast, it drags a cone of sand with it. This "sand backpack" (called added mass) makes the foot feel much heavier and creates a huge spike in force that the robot's old software couldn't explain.

The Solution: A Three-Step Detective Kit

To fix this, the team created a new "detective kit" for the robot to figure out the ground's properties while it's hopping.

1. The "Momentum Observer" (The Inertia Filter)
The robot's sensors are noisy, and the robot's own body has weight. When the robot jumps, its body moves up and down, which messes up the force readings.

The Fix: They used a mathematical tool (a Kalman filter and a momentum observer) to act like a noise-canceling headphone. It separates the force caused by the robot's own body moving from the actual force of the ground pushing back. It's like knowing exactly how much of the "jolt" you feel is from your own jump versus the ground hitting you.

2. The "Sand Physics" Model (The Backpack Theory)
They realized they needed to account for that "sand backpack."

The Fix: They updated the robot's brain to understand that when the foot accelerates (speeds up or slows down), it drags extra sand with it. They added a specific term to their math to calculate this "added mass" effect.

3. The "Smart Weighting" System (Ignoring the Chaos)
Even with the new physics, the moments when the robot hits the ground or changes direction are chaotic. The data is messy.

The Fix: Instead of trying to force the messy data to fit a perfect line, they used a "smart weighting" system. They told the computer: "When the robot is accelerating super fast (the messy part), trust the data a little less. When the robot is moving steadily, trust the data more." This allowed them to ignore the noise and find the true pattern.

The Result: From "Impact" to "Insight"

By combining these three steps, the robot could finally look at its own internal sensors and accurately guess how stiff or soft the sand was, even while hopping at high speeds.

Before: The robot guessed the sand stiffness was all over the place, depending on how fast it was jumping.
After: The robot's guess matched the "real" ground truth perfectly, regardless of speed.

Why This Matters

This isn't just about a robot hopping on a beach. This is a breakthrough for:

Planetary Exploration: Future robots sent to Mars or the Moon will need to hop over rocky, sandy, or dusty terrain at high speeds to explore quickly. They can't carry heavy external sensors; they have to "feel" the ground with their own legs.
Disaster Response: Robots that need to run through rubble or mud after an earthquake need to know instantly if the ground will hold them or swallow them.

In short, the paper teaches robots that speed changes the rules of the game. By understanding the physics of "sand backpacks" and filtering out their own body's movement, robots can finally run, hop, and explore dynamic worlds with confidence.

1. Problem Statement

Robots traversing natural, deformable terrains (e.g., sand, mud) rely on proprioceptive sensing (internal sensors like encoders and IMUs) to infer terrain mechanical properties. However, existing approaches predominantly rely on quasi-static (QS) assumptions, which model resistive forces as a function of penetration depth alone.

The Gap: These QS models fail during high-speed dynamic locomotion (e.g., hopping, running) where impact velocities and rapid stance transitions introduce significant inertial effects.
The Consequence: Under dynamic conditions, the discrepancy between QS predictions and actual ground reaction forces leads to inaccurate terrain characterization. Specifically, the paper identifies that acceleration-dependent "added-mass" effects (grain entrainment beneath the foot) and velocity-dependent drag dominate transient force responses, which QS models completely neglect. This results in poor stiffness estimation and potential locomotion failure.

2. Methodology

The authors developed a physics-based framework to characterize granular terrain using only onboard sensors, validated through controlled experiments.

A. Experimental Platform

Robot: A vertically constrained, direct-drive (gearless) hopper with a parallel five-bar linkage. The direct-drive motors provide high torque transparency, allowing force estimation from motor current.
Environment: A granular bed of 300 µm glass beads (simulating sand), prepared via a fluidized-bed technique to ensure repeatability.
Sensors:
- Proprioceptive: Motor encoders, IMUs (body and foot), Time-of-Flight (ToF) sensor.
- Ground Truth: A load cell mounted on the foot and a motion capture system (OptiTrack) for kinematic validation.
Protocol: The robot performed high-speed hopping (touchdown speeds $0.2–1.2$ m/s) with systematically varied impact speeds and leg stiffnesses. Ground-truth terrain mechanics were also measured via a linear actuator performing constant-speed intrusions.

B. State Estimation & Force Recovery

To move beyond QS assumptions, the authors implemented a two-stage estimation pipeline:

Kalman Filter (KF): Fused ToF, encoder, and IMU data to estimate body and foot kinematics (position, velocity) while filtering noise.
Momentum Observer (MO): Used the estimated kinematics to reconstruct the contact force. Unlike the QS model ( $F = J^{-T}\tau$ ), the MO solves the full dynamic equation:
$F_c = \dot{p}_f - \psi(\theta, \dot{\theta}, \dot{x}_f, \tau)$
This explicitly compensates for rigid-body inertia and gravity, isolating the external terrain reaction force from the robot's internal dynamics.

C. Physics-Based Force Decomposition

The authors analyzed the residual forces after applying the MO and found they correlated with acceleration. They modeled the granular reaction force ( $F_g$ ) as:
$F_g(z, \dot{z}, \ddot{z}) = F_d(z) + \frac{\partial m_a}{\partial z}\dot{z}^2 + m_a\ddot{z}$
Where:

$F_d(z)$ : Depth-dependent quasi-static resistance.
$\dot{z}^2$ term: Velocity-dependent quadratic drag.
$m_a\ddot{z}$ term: Acceleration-dependent added-mass effect (dominant during impact and stiffness transitions).

D. Acceleration-Aware Regression

To infer terrain stiffness ( $k$ ) from proprioceptive data, the authors introduced a Weighted Least Squares (WLS) regression.

Mechanism: Samples collected during high-acceleration events (where the added-mass model is most sensitive and variance is high) are down-weighted using a sigmoid function based on $|\ddot{z}|$ .
Goal: This prevents transient dynamic noise from skewing the estimation of the underlying static stiffness parameter.

3. Key Contributions

Identification of Dynamic Failure Modes: Demonstrated that quasi-static terrain sensing fails significantly during high-speed hopping, particularly at touchdown and controller-induced stiffness transitions, due to neglected acceleration-dependent effects.
Momentum Observer for Proprioception: Developed a momentum-observer-based estimator that successfully recovers contact forces by compensating for rigid-body inertia, bridging the gap between internal sensing and external ground truth.
Granular Physics-Informed Modeling: Validated that granular reaction forces in dynamic regimes are dominated by added-mass effects (acceleration) rather than just velocity or depth.
Robust Parameter Estimation: Introduced an acceleration-aware weighted regression that enables consistent recovery of granular stiffness parameters across varying locomotion speeds, closely matching linear-actuator ground truth.

4. Results

Force Reconstruction: The Momentum Observer (MO) significantly outperformed the quasi-static estimator. The MO closely matched load-cell ground truth throughout the stride, including the initial impact surge and force peaks during stiffness transitions, whereas the QS model failed to capture these transients.
Stiffness Estimation:
- QS Model: Produced highly nonlinear and inconsistent stiffness estimates that deviated from ground truth, especially at higher speeds.
- MO + Linear Fit: Improved accuracy by removing inertial bias but still showed residuals at high accelerations.
- MO + Acceleration-Aware WLS: Achieved the highest fidelity. The estimated stiffness ( $k_{est}$ ) tracked the ground-truth stiffness ( $k_{GT}$ ) closely across all tested touchdown speeds ( $0.22–1.22$ m/s) and leg stiffnesses.
Validation: The added-mass term derived from the physics model successfully reproduced the timing and shape of the residual forces observed during transient events.

5. Significance

Enabling Dynamic Locomotion: This work provides a foundation for legged robots to safely and efficiently traverse deformable terrains at high speeds, a capability previously limited by the reliance on slow, quasi-static sensing.
Planetary Exploration: The framework is critical for future planetary rovers (e.g., Mars, Moon) that must operate in unknown, granular environments without external instrumentation (like LiDAR or cameras) to sense ground properties in real-time.
Generalizability: The approach decouples robot dynamics from terrain mechanics, suggesting that similar "dynamics-aware" frameworks can be applied to other deformable substrates (mud, snow) and complex leg morphologies, moving robot terrain perception from static analysis to dynamic interaction.