⚡ electrical engineering

Racing a Wheeled Quadruped: Active Load Transfer Mitigation via Model Predictive Control

This paper presents a hierarchical control framework combining model predictive control and reinforcement learning to enable active roll control on a wheeled quadruped, significantly reducing lateral load transfer and improving racing performance by utilizing leg actuators as active suspension to bank into turns.

Original authors: Marla Eisman, Brian Lam, Samuel Sonnino, Francesco Borrelli

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Marla Eisman, Brian Lam, Samuel Sonnino, Francesco Borrelli

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 a robot dog that doesn't just walk or run, but actually races like a Formula 1 car. This paper describes how researchers taught a robot called the "Unitree Go2-W" to race faster and safer by teaching it to lean into turns, just like a human motorcyclist or a cyclist does.

Here is the breakdown of their work in simple terms:

The Problem: The "Flat" Robot

Most robots with wheels and legs are designed to stay perfectly upright. If you drive a flat, rigid box around a sharp corner too fast, the weight shifts to the outside wheels. This is called Load Transfer.

The Analogy: Think of a heavy suitcase on a cart. If you turn the cart sharply, the suitcase tips toward the outside. If you turn too hard, the inside wheels lift off the ground, and the whole thing might flip over.
The Robot's Issue: Because this robot has 16 motors (4 legs with 3 joints each + 4 wheels), it has a lot of moving parts. If it stays flat while racing, it hits a "speed limit" where it risks flipping or losing grip.

The Solution: The "Motorcycle" Strategy

The researchers gave the robot a new superpower: Active Roll Control.
Instead of staying flat, the robot uses its leg joints to bend asymmetrically. When it turns left, it leans its body to the left.

The Analogy: Imagine a cyclist going around a curve. They don't stay straight up; they tilt their body inward. This uses gravity to help them stick to the road and turn faster without falling.
How it works: The robot's "knees" act like a high-tech suspension system. By bending one side of the legs more than the other, the robot generates a torque that tilts its body, counteracting the force that tries to tip it over.

The Brain: A Three-Layer Team

To make this happen at racing speeds, the team built a "hierarchical" (layered) brain for the robot:

The Strategist (Offline Planning): Before the race even starts, a computer calculates the perfect path around the track (the "raceline") to get the fastest time possible.
The Pilot (MPC - Model Predictive Control): This is the robot's real-time brain. It looks at the track ahead and constantly asks, "If I go this fast and turn this hard, will I flip?" It then calculates the perfect amount of lean (roll) needed to stay safe and fast. It's like a pilot constantly adjusting the plane's banking angle.
The Muscle (RL - Reinforcement Learning): The Pilot tells the muscle what to do (e.g., "Lean 10 degrees left"), but the muscle doesn't speak "Lean." It speaks "Motor Torque." The researchers trained a Reinforcement Learning AI (like a video game character learning by trial and error) to translate the Pilot's commands into the exact movements of the 16 different motors. This AI learned how to coordinate the legs and wheels perfectly to achieve the lean.

The Results: Faster and Safer

The team tested this on an indoor track with sharp corners. They compared the robot with "Lean ON" (tilting) versus "Lean OFF" (staying flat).

Faster Laps: The leaning robot was 8.7% faster on its best lap.
More Grip: It could handle 21.3% more sideways force (lateral acceleration) before losing stability. It reached a peak of nearly 2 meters per second squared, whereas the flat robot struggled at lower speeds.
Less Risk: The "Load Transfer Ratio" (a measure of how close the robot is to tipping) dropped by 44%. The robot stayed planted on all four wheels much more effectively.

The Bottom Line

The paper proves that by treating a wheeled robot like a motorcycle—using its legs to actively lean into turns—you can make it race significantly faster and safer than if it tried to stay rigid and flat. The robot didn't just drive; it learned to "bank" its body to conquer the track.

Technical Summary: Racing a Wheeled Quadruped via Active Load Transfer Mitigation

Problem Statement
Wheeled quadrupeds offer a unique synthesis of legged adaptability and wheeled efficiency, yet their control at high speeds presents significant challenges due to complex contact dynamics and 16 degrees of freedom. In autonomous racing, vehicles operate near their dynamic limits where lateral load transfer reduces effective cornering capacity and increases rollover risk. Traditional control approaches for wheeled quadrupeds often apply legged gait dynamics or assume a rigid chassis, failing to exploit the platform's inherent degrees of freedom to actively manage body roll. While active roll control is a established strategy in motorcycles and tilting vehicles, it has not been demonstrated on wheel-legged quadrupeds for racing objectives, nor has rollover avoidance been experimentally validated for skid-steering wheeled quadrupeds in this context.

Methodology
The authors propose a hierarchical control framework integrating Model Predictive Control (MPC) and Reinforcement Learning (RL) to manage lateral load transfer on the Unitree Go2-W platform. The system operates on three levels:

Vehicle Modeling: The robot is modeled as a dynamic bicycle with an added active roll degree of freedom. The four wheels are lumped into equivalent front and rear tires, while the leg actuators function as active suspension. The model assumes a rigid body with constant geometry, where yaw is generated via differential wheel torque (skid-steering) and roll is generated by asymmetric leg extension. The dynamics include planar motion (sideslip, yaw, longitudinal speed) and decoupled roll dynamics driven by lateral acceleration.
Offline Raceline Generation: An optimal raceline is computed offline by solving a Finite-Time Optimal Control Problem (FTOCP) that minimizes lap time subject to track boundaries and vehicle dynamics. This provides a reference trajectory for position, heading, and velocity.
Online MPC Planner: A high-level MPC tracks the raceline reference. The cost function includes terms for tracking position, heading, and speed, while explicitly penalizing the Load Transfer Ratio (LTR) to drive it toward zero. The MPC outputs three high-level commands: longitudinal acceleration ( $a_x$ ), yaw moment ( $M_\psi$ ), and roll moment ( $M_\theta$ ).
Low-Level RL Policy: A reinforcement learning policy, trained in simulation using Proximal Policy Optimization (PPO) with domain randomization, maps the high-level MPC commands to 16 individual joint targets. This policy handles the highly coupled contact and motor dynamics, converting the abstract roll moment command into asymmetric leg extensions to bank the robot into turns.

Key Contributions
The paper makes three primary contributions:

MPC Formulation: A novel control formulation that utilizes active roll control to mitigate lateral load transfer, specifically tailored for the racing performance of a wheeled quadruped.
Hierarchical Architecture: A complete control stack combining offline raceline optimization, online MPC for reference tracking, and joint-level RL for execution.
Hardware Validation: Successful experimental validation of the proposed method on physical hardware, demonstrating active roll control for a racing objective on a skid-steering wheel-legged platform.

Experimental Results
Experiments were conducted on an indoor L-shaped track with the Unitree Go2-W. The proposed active roll controller ("Tilt ON") was compared against a baseline controller where the roll moment command was fixed to zero ("Tilt OFF"). Key findings include:

Load Transfer Mitigation: At matched lateral accelerations, the active roll control reduced the mean Load Transfer Ratio (LTR) by up to 44%.
Performance Gains: The active roll controller improved the fastest lap time by 8.7% (from 10.96s to 10.01s) and reduced the mean lap time by 6.4%.
Dynamic Capability: Peak lateral acceleration capability increased by 21.3%, reaching 1.98 m/s² compared to 1.64 m/s² for the baseline.
Stability: The system maintained robust high-speed stability beyond the operational range of the non-tilting baseline, though the LTR distribution showed slight over-compensation consistent with identified system gains.

Significance
The paper demonstrates that wheeled quadrupeds can transcend their rigid-body limitations by treating their legs as active suspension. By integrating active roll control into a racing framework, the system effectively counters load transfer, thereby preserving tire grip and enabling higher cornering speeds. This work bridges the gap between vehicle dynamics strategies used in tilting vehicles and the control of legged robots, proving that active body lean can be a viable and effective strategy for enhancing the dynamic performance of wheel-legged platforms in high-speed autonomous racing.

The Problem: The "Flat" Robot

The Solution: The "Motorcycle" Strategy

The Brain: A Three-Layer Team

The Results: Faster and Safer

The Bottom Line

More like this