Ground Reaction Force-Aware Model Predictive Control for Underactuated Upright Physical Assistance

Darren Biskup, Jeonghan Yu, Zibo Yu, Sunil K. AgrawalarXiv 2026 · DOI · arXiv

TL;DR

This paper introduces GRF-MPC, a model predictive control framework for rehabilitation robots that incorporates real-time ground reaction force (GRF) measurements into a human centroidal dynamics model. By accounting for how a user interacts with the ground, the system provides more efficient, transparent, and effective physical assistance—outperforming traditional impedance control even when using half the number of actuators.

What Problem Does This Paper Address?

Most current rehabilitation robots, such as exoskeletons or cable-driven stand trainers, use "reactive" tracking controllers. These systems typically try to force a user to follow a pre-defined path; if the user deviates, the robot pushes back.

The fundamental flaw in this approach is that it ignores the human-ground interaction. When a person stands or moves, they generate stability and propulsion through forces at their feet. Without "knowing" these ground reaction forces, a robot might interpret a user's natural, volitional movement as a tracking error and fight against it. This leads to assistance that feels "opaque" or restrictive, often requiring the user to compensate for the robot's own corrective forces. The authors aim to bridge this gap by treating the human-ground interaction as a vital, live observation for the control loop.

Methods

The researchers developed the Ground Reaction Force-Aware Model Predictive Control (GRF-MPC) framework and tested it on the Robotic Upright Stand Trainer (RobUST), a cable-driven system.

  • Centroidal Dynamics Model: Instead of a complex multi-body model, they used a Single Rigid Body (SRB) approximation. This focuses on the movement of the human's Center of Mass (CoM) and angular momentum.
  • Live GRF Integration: The system uses floor-mounted force plates to measure net GRF and Center of Pressure (CoP) in real-time. These measurements are fed directly into the dynamics model, transforming an otherwise unobservable human intent into a predictable physical state.
  • Convex Optimization: The controller formulates a Quadratic Program (QP) to optimize cable tensions over a 0.5-second look-ahead window. This formulation natively handles hardware constraints like cable tension limits and underactuation.
  • Hierarchical Control: A high-level MPC planner runs at 100 Hz to calculate optimal tensions, while a low-level PID loop runs at 1 kHz to ensure those tensions are accurately tracked by the motors.
Fig. 1. (A) Schematic of the Robotic Upright Stand Trainer (RobUST) system highlighting motion tracking, cable actuation, and CoP/GRF measurement. (B) Photograph of the experimental setup showing the user equipped with Vive trackers, standing on force plates, and receiving visual feedback. (C) Real-time graphical interface displaying the user’s inertia ellipsoid, target and predicted positions, and force vectors.

Datasets

The framework was validated through a human-subject study involving six healthy participants (3 male, 3 female). The evaluation consisted of nearly 300 trials across three distinct tasks designed to test different aspects of the controller:

  1. Stable Standing + Perturbation: Testing the ability to arrest sudden momentum from an unexpected 150 N pulse.
  2. Deep Lunge: Testing dynamic tracking precision during a large, 20 cm vertical drop.
  3. Permitted Workspace Mapping: Measuring "transparency" by seeing how far a user could voluntarily lean before the robot forced a return.

Key Insights and Ideas

  • GRF as the "Enabling Observation": In legged robotics, GRF is usually a control output. For human assistance, it's an input generated by the user. Measuring it is the "secret sauce" that allows the robot to distinguish between a disturbance (like a push) and a user's own intent to move.
  • Algorithmic Intelligence vs. Hardware Complexity: One of the most striking findings was that the underactuated 4-cable GRF-MPC outperformed a fully-actuated 8-cable impedance baseline. This suggests that better software models can reduce the need for heavy, complex hardware.
  • Proactive vs. Reactive: Because MPC looks ahead, it can proactively coordinate tensions to track a reference more efficiently, rather than just reacting to errors after they occur.

Key Results

The GRF-MPC showed significant improvements over the state-of-the-art impedance control baseline across all metrics:

  • Superior Disturbance Rejection: Reduced peak planar CoM error by 19.6% during sudden perturbations.
  • Better Tracking: Improved trajectory tracking precision by 24.1%.
  • Higher Efficiency: Required 28.7% less net mechanical work from the robot to achieve these results.
  • Increased Transparency: Permitted a 45.9% larger voluntary workspace, meaning the robot was less "intrusive" during natural movement.
  • Real-time Performance: The underlying QP solver consistently ran in under 1 ms, making it highly suitable for high-frequency robotic control.
Quantitative results across the three experimental tasks, showing the superiority of MPC in error reduction, work efficiency, and workspace preservation.

Limitations and Future Work

While promising, the authors note several areas for improvement:

  • Sensing Constraints: The current system relies on bulky, floor-mounted force plates. Future versions would need wearable sensors (like pressure-sensing insoles) or GRF estimation algorithms to work outside of a lab.
  • Model Simplifications: The Single Rigid Body model assumes a constant inertia tensor. While sufficient for upright tasks, more complex movements might require accounting for how a user's pose changes their inertia.
  • Clinical Validation: This study was limited to healthy young adults. The next critical step is testing the system with individuals with neurological or musculoskeletal impairments (e.g., stroke or spinal cord injury).
  • Self-Paced Pacing: The authors suggest moving toward Model Predictive Contouring Control (MPCC), which would allow users to progress through a movement at their own pace rather than following a strict, time-based trajectory.

Summary updated

Abstract

Many rehabilitation robots execute prescribed motions using reactive tracking controllers that do not explicitly incorporate the user's instantaneous human-ground interaction. This paper presents Ground Reaction Force-Aware Model Predictive Control (GRF-MPC), a framework for upright physical assistance. By integrating live GRF and center of pressure (CoP) measurements directly into a human centroidal dynamics model, GRF-MPC formulates a real-time quadratic program (QP) that optimizes assistive forces and moments over a receding horizon. We deployed GRF-MPC on the Robotic Upright Stand Trainer (RobUST) in an underactuated 4-cable configuration and evaluated it against a fully actuated 8-cable Cartesian impedance baseline in nearly 300 trials across six healthy participants. Across tasks spanning impulsive perturbations, lunge trajectories, and voluntary limits of stability, GRF-MPC reduced peak planar center-of-mass (CoM) error by 19.6% during perturbations, lowered trajectory tracking error by 24.1%, required 28.7% less net mechanical work, and permitted a 45.9% greater normalized voluntary excursion area. The underlying QP solves in under 1 ms, comfortably enabling 100 Hz replanning and demonstrating that incorporating live human-ground interaction into predictive control yields stronger, more transparent physical assistance with half the actuators.
BibTeX
@misc{biskup2026ground,
  author = {Darren Biskup and Jeonghan Yu and Zibo Yu and Sunil K. Agrawal},
  title = {{Ground Reaction Force-Aware Model Predictive Control for Underactuated Upright Physical Assistance}},
  year = {2026},
  eprint = {2611.13204},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13204},
}
Figure