💻 computer science

RK-MPC: Residual Koopman Model Predictive Control for Quadruped Locomotion in Offroad Environments

This paper introduces RK-MPC, a real-time, data-driven model predictive control framework that combines a nominal template model with a learned linear residual predictor to enable robust, high-frequency quadruped locomotion across diverse and unstructured off-road terrains.

Original authors: Sriram S. K. S. Narayanan, Umesh Vaidya

Published 2026-04-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sriram S. K. S. Narayanan, Umesh Vaidya

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 teaching a dog to run through a forest.

If you just tell the dog, "Run straight," it might trip over a root, slip on a patch of mud, or get confused by a sudden hill. If you try to write a perfect physics textbook describing exactly how every paw interacts with every type of leaf, rock, and snowflake, the book would be thousands of pages long, and the dog would be too slow to read it before falling over.

This paper presents a solution called RK-MPC (Residual Koopman Model Predictive Control). Think of it as giving the robot dog a smart, two-part brain that combines a "rulebook" with a "street-smart experience."

Here is how it works, broken down into simple concepts:

1. The "Rulebook" (The Nominal Model)

First, the robot has a basic understanding of how it should move. This is like a physics textbook. It knows that if it pushes its leg forward, it should move forward. It knows gravity pulls it down.

The Problem: This rulebook is great for a flat, smooth gym floor. But in the real world (off-road), the ground is messy. The rulebook doesn't know about slippery ice, deep snow, or loose gravel. If the robot relies only on this rulebook, it will make mistakes and fall.

2. The "Street Smarts" (The Residual Learner)

This is the magic part. Instead of trying to rewrite the entire physics textbook to account for every possible rock and puddle, the robot learns a small "correction note."

The Analogy: Imagine you are driving a car. The rulebook says, "If I turn the wheel 10 degrees, the car turns 10 degrees." But you know from experience that on wet roads, you need to turn 12 degrees to get the same result.
The "Residual" part of this system is like a co-pilot who whispers, "Hey, the road is slippery today, so turn a little more than the rulebook says."
It learns this "whisper" by watching the robot move and noticing the difference between what the rulebook predicted and what actually happened. It learns a simple, fast pattern to fix the mistakes.

3. The "Koopman" Trick (Making it Linear)

Usually, learning these corrections is hard because the world is chaotic and non-linear (small changes can cause big, unpredictable results).

The Metaphor: Imagine trying to draw a perfect circle on a piece of crumpled paper. It's hard. But if you could magically "lift" that paper into a higher dimension where it becomes flat again, drawing the circle becomes easy.
The Koopman method is that magic trick. It takes the messy, complex data of the robot moving on rough terrain and "lifts" it into a mathematical space where the rules become simple and straight lines (linear). This allows the robot to do complex math very, very quickly.

4. The "Predictive Planner" (MPC)

Now, the robot doesn't just react; it plans ahead.

The Analogy: Think of a chess player. They don't just move a piece based on what's happening right now; they look 5 or 10 moves ahead to see if they will get trapped.
The MPC (Model Predictive Control) is the chess player. Every split second, it runs a simulation: "If I step here, will I slip? If I step there, will I hit a rock?"
It uses the Rulebook + the Street Smarts (Residual) to make these predictions. Because the math is kept simple (thanks to the Koopman trick), it can do this planning 500 times a second—fast enough to keep the robot balanced while running at full speed.

Why is this a big deal?

The researchers tested this on a real robot (a Unitree Go1, which looks like a small dog) in some very tough conditions:

Grass and Gravel: The robot didn't slip.
Snow and Ice: The robot adjusted its steps to not slide away.
Blind Locomotion: The robot didn't use cameras to see the ground. It just felt the ground with its feet and adjusted instantly.
Different Gaits: They taught it to "trot" (like a horse), but it could also "crawl" (slow and steady) without needing to be retrained.

The Bottom Line

Previous methods were like trying to memorize every possible path through a forest (too slow) or just guessing based on trial and error (too dangerous).

RK-MPC is the sweet spot. It keeps the reliable physics rules but adds a tiny, super-fast "learning layer" that fixes the mistakes in real-time. It's like giving a robot a PhD in physics and a lifetime of experience walking through mud, all while keeping its brain light enough to run on a small computer attached to its back.

This means we can finally send robots into disaster zones, forests, or snowy mountains to do work, confident that they won't trip over a single rock.

1. Problem Statement

Quadruped robots operating in unstructured, off-road environments face significant challenges in maintaining stable locomotion due to:

Model Mismatch: Classical physics-based Model Predictive Control (MPC) relies on reduced-order dynamics (e.g., Single Rigid Body, SRB) which often fail to capture complex interactions like terrain compliance, variable friction, and contact uncertainty.
Data-Driven Limitations: Purely data-driven approaches (e.g., Deep Reinforcement Learning) often lack interpretability, safety guarantees, and the ability to handle hard constraints (friction, contact) explicitly.
Koopman Operator Challenges: While Koopman-based methods offer a way to linearize nonlinear dynamics for control, applying them directly to quadrupeds has proven difficult.
- Monomial EDMD: Standard polynomial lifting often leads to unstable multi-step predictions and "blow-up" over time.
- SE(3)-Structured EDMD: While more geometrically consistent, these methods suffer from rotation drift. Small errors in attitude estimation compound over the prediction horizon, causing the internal rotation matrix to drift off the $SO(3)$ manifold, leading to controller instability.
Real-Time Constraints: High-fidelity learning models often introduce computational overhead that prevents real-time execution on embedded hardware.

The core problem is how to improve prediction fidelity for quadruped locomotion in unstructured environments while preserving the convex structure, constraint handling, and real-time solvability of optimization-based control.

2. Methodology: Residual Koopman MPC (RK-MPC)

The authors propose RK-MPC, a hybrid framework that combines a physics-based nominal model with a data-driven residual correction learned via Koopman operator theory.

A. Core Architecture

The system uses a hierarchical control stack:

Nominal Model: A linearized Single Rigid Body (SRB) model serves as the "backbone." It captures the dominant dynamics but is simplified (e.g., neglecting gyroscopic coupling and assuming small roll/pitch angles).
Residual Learning: Instead of learning the full state dynamics, the system learns a linear residual predictor for the modeling error ( $e_k = x_k - x^{nom}_k$ $e_{k} = x_{k} - x_{k}^{n o m}$ ).
- The residual is defined specifically over the centroidal twist (linear and angular velocity errors).
- Koopman Lifting: The residual error is lifted into a higher-dimensional space using a dictionary of observables (specifically, degree-2 monomials).
- EDMD with Control (EDMDc): The lifted residual dynamics are approximated as a linear time-invariant system: $z_{k+1} = A_{res}z_k + B_{res}u_k$ .
Correction: The final prediction is the sum of the nominal SRB prediction and the learned residual correction:
$\hat{x}_{k+1} = x^{nom}_{k+1} + C_{res}(A_{res}z_k + B_{res}u_k)$

B. Convex MPC Formulation

The learned residual model is embedded directly into a convex Quadratic Program (QP):

Objective: Minimize tracking error and control effort.
Dynamics: The linearized SRB dynamics are augmented with the linear residual term.
Constraints: The formulation retains standard legged constraints, including:
- Friction cones (approximated as pyramids).
- Vertical force bounds.
- Contact schedules (swing vs. stance phases).
Real-Time Performance: Because the residual model is linear and the nominal model is linear, the entire optimization remains a convex QP, solvable efficiently on embedded hardware.

C. Theoretical Guarantees

The paper provides a formal theorem proving that if the nominal dynamics are Lipschitz continuous and the residual approximation error is bounded, the multi-step prediction error of the RK-MPC remains bounded. This ensures that the learned correction does not destabilize the system over long horizons.

3. Key Contributions

Residual Koopman Modeling: A novel framework that learns a compact linear residual correction on top of a physics-based template. This approach avoids the instability of full-state lifting while correcting systematic model mismatches (terrain, friction, contact variability).
Convex RK-MPC Formulation: The integration of the learned residual into a receding-horizon controller that runs at 500 Hz on embedded hardware (NVIDIA Jetson Xavier NX), maintaining the computational efficiency of convex optimization.
Robustness to Observable Choice: The method significantly reduces sensitivity to the choice of observables compared to full-state EDMD. By focusing only on the residual, it avoids the "rotation drift" issues seen in SE(3)-based lifts.
Hardware Validation: Extensive validation on a Unitree Go1 robot in diverse, unstructured environments (grass, gravel, snow, ice) and under external disturbances (pushes, debris), demonstrating "blind" locomotion without exteroceptive sensing.

4. Experimental Results

The authors compared RK-MPC against three baselines:

SRB-MPC: Standard physics-based MPC (non-convex or linearized).
EDMD-Mono: Full-state Koopman with monomial dictionary.
SE3-KMPC: Full-state Koopman with SE(3)-structured dictionary.

Key Findings:

Prediction Accuracy: In open-loop tests, RK-MPC achieved the lowest Root Mean Square Error (RMSE) for velocity and angular rate predictions. Monomial EDMD diverged rapidly, and SE3-KMPC suffered from attitude drift.
Closed-Loop Stability:
- SE3-KMPC failed within seconds due to rotation manifold drift.
- RK-MPC successfully completed 10-lap circular tracking tasks and maintained stability comparable to the high-fidelity SRB-MPC baseline.
Computational Efficiency: RK-MPC had the fastest solve time (0.89 ms on average), outperforming both the non-convex SRB-MPC (1.57 ms) and SE3-KMPC (1.57 ms).
Off-Road Robustness:
- Debris/Slip: The robot successfully traversed debris fields and recovered from external pushes without falling.
- Terrain Generalization: The controller, trained on trot gait data, successfully generalized to a crawl gait without retraining.
- Unstructured Terrains: The robot maintained stable velocity tracking on grass, gravel, snow, and ice, adapting to varying friction coefficients and terrain compliance.

5. Significance

This work presents a practical pathway for deploying data-driven predictive control on legged robots in real-world scenarios.

Bridging the Gap: It effectively bridges the gap between the robustness of physics-based models and the adaptability of learning-based methods.
Safety and Structure: By keeping the core optimization convex and physics-constrained, it ensures safety guarantees that are often missing in pure learning approaches.
Hardware Viability: The demonstration of 500 Hz control on an embedded processor proves that sophisticated, data-augmented control is feasible for real-time, agile quadruped locomotion.
Scalability: The residual approach suggests a scalable method for handling complex dynamics without the computational explosion of full-state lifting, making it applicable to other high-dimensional robotic systems.

In summary, RK-MPC offers a "best of both worlds" solution: the interpretability and constraint handling of model-based control, enhanced by the adaptability of data-driven learning, specifically tailored for the challenges of off-road quadruped locomotion.