⚡ electrical engineering

Impedance MPC for Physical Human-Robot Interaction: Predictive Disturbance Rejection with Joint-Limit Safety

This paper presents a two-layer Impedance MPC framework that combines analytical dynamic linearization with a high-frequency convex QP and an augmented Kalman filter to achieve sub-0.05 mm steady-state error under sustained human forces while ensuring joint-limit safety, significantly outperforming classical impedance control on a 7-DOF Franka FR3 robot.

Original authors: Yongyan Cao, Jinshan Tang

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Yongyan Cao, Jinshan Tang

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ 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 arm working alongside a human in a shared space, like a factory floor or a hospital room. The robot needs to be precise enough to follow a specific path (like drawing a perfect circle) but also soft and compliant enough that if a human accidentally bumps into it, the robot yields safely instead of fighting back or breaking something.

This paper presents a new "brain" for these robots, called Impedance MPC, that solves a classic problem: How do you stay on track without getting pushed off course?

Here is the breakdown of how it works, using everyday analogies.

The Problem: The "Stiff Spring" vs. The "Sticky Glue"

Traditional robot control acts like a stiff spring.

The Scenario: You push the robot arm with a steady 15-pound force.
The Result: The spring stretches. The robot moves away from its target path and stays there. To get it back, you have to push it back, or make the spring incredibly stiff.
The Catch: If you make the spring too stiff to stop the movement, the robot becomes dangerous. If a human bumps it, the robot fights back hard, which defeats the purpose of a "collaborative" robot.
The Old Fix: Engineers tried adding a "memory" (integral action) to the spring to remember the push and correct it. But this was like trying to balance a broom on your hand while running; it was unstable and slow.

The Solution: A Two-Layer "Smart Pilot"

The authors built a two-layer system that acts like a highly skilled pilot who can predict the wind before it hits the plane.

Layer 1: The "Gravity Canceler" (The Pre-Flight Check)

Robots are heavy and their joints interact in complex ways (like a gymnast swinging on a bar).

The Analogy: Imagine trying to steer a heavy boat in a river with strong currents. It's hard to predict where you'll go.
What the Robot Does: Layer 1 instantly calculates and cancels out the robot's own weight, the friction of its joints, and the "swing" of its movement.
The Result: It turns the complex, wobbly robot arm into a simple, predictable object—like a frictionless puck sliding on ice. This simplifies the math so much that the computer can solve the next problem incredibly fast.

Layer 2: The "Crystal Ball" (The Model Predictive Controller)

Now that the robot is a simple "ice puck," Layer 2 takes over. This is the Model Predictive Controller (MPC).

The Analogy: Imagine driving a car while looking 10 seconds ahead. If you see a pothole coming, you steer slightly now so you don't hit it later.
What the Robot Does: Instead of just reacting to a push after it happens, this layer solves a complex math puzzle 100 times a second. It asks: "If I apply a tiny force now, where will I be in 0.1 seconds? In 0.2 seconds?"
The Magic: It finds the perfect sequence of tiny moves to stay on the path, even if a human is pushing it. It respects "safety zones" (joint limits) just like a driver respects lane markings.

The Secret Sauce: The "Noise-Canceling Headphones"

Even with the crystal ball, a steady human push (like a constant 15N force) can still trick the system.

The Innovation: The authors added a Kalman Filter that acts like noise-canceling headphones.
How it Works: The system constantly listens to the "noise" (the human's push). It estimates exactly how hard the human is pushing and adds an equal and opposite force to cancel it out perfectly.
The Result: The robot doesn't just "give in" to the push; it actively fights the push to stay exactly on its original path.

The Results: From "Wobbly" to "Laser-Sharp"

The team tested this on a 7-armed robot (Franka FR3). Here is what happened when a human pushed it with a steady 15-pound force:

Old Robots (Classical Impedance): The robot moved 44.8 millimeters (almost 2 inches) off its path and stayed there.
New Robot (Impedance MPC + Kalman): The robot moved less than 0.05 millimeters off its path.
- Translation: That is a 800-fold improvement. The new robot is so precise it's almost as if the human never pushed it at all.

Safety Features

The system also has a built-in "guardian angel" for the robot's joints:

The Analogy: Imagine a runner who knows they are approaching a wall. As they get close, they instinctively slow down and steer away, even if the coach (the main task) is yelling "Run straight!"
The Tech: The robot uses a "barrier potential" to sense when a joint is about to hit its physical limit. It automatically redirects the movement into a safe direction without breaking the main task.

Summary

This paper introduces a robot control system that:

Simplifies the physics of the robot to make it easy to calculate.
Predicts the future to steer around disturbances before they cause big errors.
Listens to the environment to cancel out constant pushes perfectly.
Keeps the robot safe by never letting its joints hit their limits.

The result is a robot that can work side-by-side with humans, staying perfectly on its task even when being pushed, without becoming dangerous or stiff.

Technical Summary: Impedance MPC for Physical Human–Robot Interaction

Problem Statement
Physical human–robot interaction (pHRI) requires robots to simultaneously maintain high trajectory accuracy and ensure compliant safety during unplanned contact. Classical impedance control, while effective for safety, suffers from a fundamental structural limitation: under a sustained human force ( $F_h$ ), it exhibits a non-zero steady-state position error defined by $e_\infty = K_d^{-1}F_h$ . For example, a 15 N push against a 300 N/m stiffness results in a 50 mm deflection. While increasing stiffness ( $K_d$ ) reduces this error, it violates safety constraints by increasing contact forces. Conversely, adding integral action to eliminate steady-state error is severely limited by stability requirements, restricting convergence speed. Existing Model Predictive Control (MPC) approaches for pHRI often retain full nonlinear plant dynamics, limiting control rates, or fail to provide formal guarantees for zero steady-state error under persistent disturbances.

Methodology
The authors propose a two-layer Impedance MPC architecture designed to resolve the tension between trajectory accuracy and compliant safety.

Layer 1: Analytical Nonlinear Inversion
The first layer employs a feedforward control strategy based on operational-space decomposition. It analytically cancels gravity, Coriolis, and task-space inertia terms. This transforms the residual plant dynamics into a configuration-independent linear double integrator.
- The error dynamics are reduced to $\ddot{e} = -\Lambda^{-1}(q) F_{mpc} + d(t)$ , where $d(t)$ represents aggregated disturbances (human force, friction, model errors).
- Crucially, this linearization results in a constant state-transition matrix ( $A_d$ ) for the discrete-time system, regardless of the robot's configuration. This allows the "free-response" matrix to be precomputed once, significantly reducing online computational load.
Layer 2: Convex QP with Disturbance Augmentation
The second layer solves a receding-horizon optimization problem at 100 Hz (with a 1 kHz inner loop for feedforward updates).
- Disturbance Estimation: An augmented Kalman filter estimates a persistent disturbance state ( $\hat{d}$ ) modeled as a random walk. This state is integrated into the MPC prediction model, enabling offset-free tracking (zero steady-state error) for constant or asymptotically constant disturbances.
- Optimization: The controller solves a 30-variable convex Quadratic Program (QP) subject to box constraints on the control force ( $F_{mpc}$ ). The cost function minimizes tracking error over a horizon while respecting joint limits.
- Safety Mechanisms: Joint-limit safety is enforced via a dual-barrier approach: a null-space inverse-barrier potential and a task-space workspace projection. These mechanisms prevent joint-limit violations without degrading task tracking in safe regions.

Key Contributions

Two-Layer Constant- $A_d$ Architecture: By analytically canceling nonlinearities, the authors reduce the plant to a linear time-varying system with a constant state matrix. This enables a 30-variable QP to be solved in under 1 ms, making real-time implementation feasible.
Analytical Connection to Classical Impedance: The paper proves (Theorem 1) that in the unconstrained, disturbance-free limit, the proposed MPC is mathematically equivalent to a classical task-space impedance controller. The realized stiffness and damping are determined by the LQR gains derived from the cost weights. This positions the method as an extension of classical impedance that adds constraint handling and offset-free rejection, rather than a replacement.
Formal Zero Steady-State Error Guarantee: Through Kalman disturbance augmentation, the system achieves an Internal Model Principle-based guarantee of zero steady-state error for persistent disturbances, extending prior offset-free MPC theory to torque-controlled manipulators.
Dual-Barrier Joint-Limit Safety: The integration of null-space potentials and workspace projections empirically prevents joint-limit violations across the tested workspace, including boundary stress tests.

Experimental Results
Experiments were conducted on a 7-DOF Franka FR3 robot using MuJoCo simulation. The proposed method (Impedance MPC with Kalman augmentation) was compared against classical impedance, admittance control, and PI impedance.

Steady-State Error: Under a sustained 15 N force, the proposed method achieved a steady-state error of **<0.05 mm**, compared to **44.8 mm** for classical impedance (an >800-fold reduction).
Transient Performance: In circular trajectory tracking, the MPC alone (without Kalman) reduced the contact-window RMS error by 19× compared to classical impedance. The addition of the Kalman filter eliminated the steady-state error entirely.
Robustness: The system demonstrated graceful robustness to measurement noise (steady-state error remained <1 mm even with 5 mm noise) and inertial model mismatch (tracking RMSE remained invariant up to 30% inertia error when the Kalman filter was active).
Safety: In boundary tests where classical impedance violated joint limits, the MPC maintained a safe margin (minimum joint margin of 0.084 vs. 0.048 for impedance).

Significance and Claims
The paper claims that the proposed architecture successfully unifies the competing demands of pHRI: high-precision trajectory tracking and compliant safety. The primary significance lies in the structural insight that analytical feedforward cancellation enables a constant linear structure, which in turn permits:

Precomputation: The free-response matrix is computed once, decoupling the QP complexity from the robot's configuration.
Formal Guarantees: The linear structure allows for rigorous proofs of impedance equivalence and offset-free tracking.
Real-Time Feasibility: The approach achieves high control rates (100 Hz QP, 1 kHz feedforward) on standard hardware, overcoming the computational bottlenecks of previous nonlinear MPC approaches.

The authors conclude that this method provides a robust, theoretically grounded solution for pHRI that eliminates the steady-state error inherent in classical impedance control while maintaining safety constraints, with future work planned for physical hardware validation and energy-tank passivity augmentation.