⚡ electrical engineering

An Integrated Real-Time Testbed for Modeling, Control, and Experimental Evaluation of a Single-Legged Hopper

This paper presents an integrated real-time testbed within MATLAB/Simulink that enables rapid prototyping and experimental evaluation of a two-link single-legged hopper, successfully demonstrating forward hopping at speeds up to 1.59 m/s while validating a phase-dependent controller and documenting key hardware-model discrepancies.

Original authors: John Bennett, Siyuan Xing

Published 2026-07-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: John Bennett, Siyuan Xing

Original paper licensed under CC BY 4.0 (https://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

The Science of Bouncing: Why Robots Need to Learn to Jump

Imagine trying to teach a robot to walk. It sounds simple, but the ground is a tricky partner. When a robot's foot hits the floor, the rules of physics change instantly. One moment, the robot is flying through the air like a projectile; the next, it is slamming into the ground, compressing, and pushing off again. This back-and-forth between flying and touching down is called a "hybrid system." It's like a dancer who has to switch between a graceful leap and a sudden, heavy stomp without missing a beat. If the timing is off, the robot trips, falls, or wastes all its energy.

Scientists have long studied how animals and simple machines handle this dance. They use models—simplified mathematical maps of how things move—to predict what will happen. But there's a catch: a model is just a drawing on a computer. To know if it really works, you have to build the robot and make it jump. The problem is that building a robot that can jump, writing the code to control it, and testing it safely is a massive headache. It usually requires a team of experts in mechanics, electronics, and computer science all working together. This paper tackles that headache by building a "playground" where researchers can design, test, and tweak their jumping robots all in one place, using tools that look like the software engineers already know.

The Paper: Building a Robot Jumping Playground

This paper introduces a new, integrated system designed to make testing single-legged hopping robots much easier. Think of it as a "rapid-control-prototyping" testbed. Instead of building a robot, writing code, and then realizing the code doesn't talk to the motors correctly, this system lets researchers design the robot's brain and body inside a single software environment called MATLAB/Simulink. They can then send that brain directly to a real robot and watch it jump, all while the computer records every tiny movement and electrical signal in real-time.

The robot itself is a "hopper" with two legs (or rather, two links) and a hip, mounted on a giant circular arm. This arm acts like a safety tether, letting the robot hop forward in a circle without falling over or running away. The researchers used this setup to test a specific way of controlling the robot: splitting the jump into two distinct phases. First, there's the flight phase, where the robot is in the air. Here, the controller plans a smooth path for the legs to follow, like a gymnast planning their flip. Then, there's the stance phase, where the foot hits the ground. Here, the robot switches tactics, using a mix of "feedforward" (guessing how hard to push based on the plan) and "impedance control" (acting like a spring to absorb the shock and push back).

The team found that in their computer simulations, the robot could learn to hop repeatedly in about four seconds, settling into a steady rhythm. They also discovered something interesting about how to make the robot go faster. They tested two "knobs" they could turn: one that changed how fast the robot touched down (called $\alpha$ ) and another that changed where the robot planned to lift its foot off the ground (called $P_x$ ). The simulations suggested that turning the $P_x$ knob had a much bigger effect on speed than the $\alpha$ knob.

When they moved from the computer to the real, physical robot, the results were a mix of success and reality checks. The robot successfully performed repeated forward hops, reaching a top average speed of 1.59 meters per second. It followed the same basic rhythm as the simulation, with short stances and longer flights. However, the real robot wasn't perfect. The computer model predicted a smooth transition, but the real robot had a tiny delay of about 0.03 seconds between sensing the ground and starting to push off. Also, while the simulation showed a clear trend that changing the lift-off position ( $P_x$ ) made the robot faster, the real-world data was a bit messy. The researchers could confirm that changing $P_x$ did increase speed, but they couldn't clearly see the same effect from the touch-down speed knob ( $\alpha$ ) because the data wasn't quite precise enough to show a clear pattern.

The paper concludes that while the system works and is a great tool for teaching and research, there are still gaps between the math and the metal. The real robot's foot is more flexible than the model assumed, and the timing of the "touchdown" signal is slightly off. The authors suggest that to get even better results, future work needs to focus on better sensors for measuring exactly how hard the foot hits the ground and refining the timing of the switches between flying and landing. They didn't claim to have solved the problem of perfect jumping robots, but they did build a very useful, transparent workshop to help others figure it out.

Technical Summary: An Integrated Real-Time Testbed for Modeling, Control, and Experimental Evaluation of a Single-Legged Hopper

Problem Statement
Dynamic legged locomotion involves hybrid dynamical systems where governing equations and constraints change abruptly at touchdown and lift-off. While the single-legged hopper serves as a canonical platform for studying these mechanics, reproducing a complete experimental stack remains demanding. Researchers and students trained primarily in mechanical systems often face a high integration burden when attempting to bridge modeling assumptions, control logic, and physical hardware. Existing platforms vary widely in morphology and control implementation, making it difficult to distinguish discrepancies arising from locomotion principles versus those caused by specific hardware realizations. There is a need for an inspectable, end-to-end workflow that unifies hybrid modeling, phase-dependent control, and real-time physical testing within a common environment.

Methodology
The authors developed an integrated testbed using MATLAB/Simulink and a Speedgoat real-time target to control a two-link robotic leg mounted on a circular planarizer. The methodology is structured around four core components:

Hybrid Mathematical Modeling: The system dynamics are derived separately for the flight and stance phases using the Euler-Lagrange method.
- Flight Phase: Modeled as a four-degree-of-freedom system (hip translation and link rotation) with projectile motion.
- Stance Phase: Modeled with ground reaction forces (GRF) represented by spring-damper elements to simulate foot compliance and collision.
- Phase Transitions: The two phases are connected via momentum-consistent state transition maps. Touchdown is modeled as an instantaneous impact conserving linear and angular momentum, while lift-off is triggered by a kinematic condition (vertical spring elongation with positive velocity).
Phase-Dependent Control: A controller is implemented that switches logic based on the locomotion phase:
- Flight Control: Utilizes a trajectory planner generating cubic polynomial trajectories for joint angles, bounded by lift-off and touchdown conditions. A PD controller tracks these trajectories.
- Stance Control: Combines feedforward torque (derived from a desired half-sine GRF profile) with impedance control (PD control) to regulate joint stiffness and damping. The controller aims to achieve a specific lift-off direction and velocity.
Real-Time Implementation: The control architecture is deployed on a Speedgoat target via a transparent Simulink structure. The system uses a Stateflow state machine to manage operating modes (zeroing, initialization, hopping, shutdown) and phase transitions. Communication with the actuator drives and encoders occurs over CAN bus, allowing for synchronized data acquisition and parameter tuning between trials.
Experimental Setup: The physical platform consists of a two-link leg with a non-rotating hip assembly, mounted on a rigid aluminum boom to permit repeated forward hopping in a compact workspace. The setup includes joint encoders, motor current sensors, and planarizer encoders to reconstruct hip kinematics.

Key Contributions
The paper documents a complete workflow from parameter identification and hybrid modeling to numerical simulation and real-time physical testing. Specific contributions include:

Integrated Workflow: A modular Simulink implementation that exposes CAN communication, phase detection, trajectory generation, and stance control, allowing parameters and signals to be accessed directly from the host computer.
Hybrid Dynamics Implementation: A demonstration of how to connect separate flight and stance models via momentum-consistent transition maps within a real-time environment.
Experimental Validation: The execution of repeated forward hopping on physical hardware, recording phase-level signals to evaluate command execution and transition timing.
Sensitivity Analysis: Preliminary numerical and hardware sweeps of the touchdown-velocity scaling factor ( $\alpha$ ) and the lift-off position parameter ( $P_x$ ) to analyze their effect on average forward speed.

Results

Simulation: Numerical simulations demonstrated that the system approaches a repeatable hopping response within approximately four seconds. The flight controller closely tracked joint reference trajectories, while the stance controller showed that not all joint states reached their terminal targets before lift-off due to the simultaneous action of impedance and feedforward torques. Parameter sweeps predicted a monotonic increase in average forward speed with increases in both $\alpha$ and $P_x$ , with $P_x$ showing a larger sensitivity.
Hardware Experiments: The physical leg successfully executed repeated forward hops, reaching a maximum measured average speed of 1.59 m/s. The system correctly executed the intended four-state phase sequence (flight, touchdown detection, stance, lift-off detection).
- Phase Execution: A timing discrepancy was observed where the electrical impact cue (motor current derivative) preceded stance activation by approximately 0.03 seconds.
- Tracking: While flight-phase joint trajectories qualitatively matched simulation, stance-phase errors were larger and followed a different evolution on hardware compared to the model.
- Sensitivity: Hardware sweeps confirmed the positive trend between average speed and the lift-off position parameter ( $P_x$ ). However, the available data for the touchdown-velocity scale ( $\alpha$ ) were insufficient to resolve the predicted positive sensitivity trend.
- Actuator Loading: Peak motor currents reached approximately 30 A, remaining below the documented 40 A rating, with no evidence of hard current clipping.

Significance and Claims
The paper claims to establish an end-to-end rapid-control-prototyping environment that reduces the embedded-software burden for users trained in mechanical systems and control. By integrating modeling, control, and experimental evaluation into a single MATLAB/Simulink-Speedgoat workflow, the platform allows for the direct examination of modeling assumptions, control logic, and measured behavior. The work highlights the principal sensing, timing, and model-hardware discrepancies encountered during implementation, specifically identifying force sensing, planarizer identification, and transition timing as priorities for further development. The authors position this work as a tool for legged-robotics research and hands-on instruction, emphasizing platform transparency and operability over assessed learning outcomes.

The Science of Bouncing: Why Robots Need to Learn to Jump

The Paper: Building a Robot Jumping Playground

More like this