Multi-Contact Feedback MPC | Intelligent Robotic Systems Lab (IRSL)

Model predictive control
Physical HRI

Real-time model predictive control uses multi-contact feedback to regulate contact forces during interactive tasks, including unexpected pushes and unknown loads.

IEEE ICRA · 2024

Safe Predictive Control under Unexpected Contacts

Interactive Tasks with Unexpected Contacts

A robot performing a contact task may also be pushed by a person or encounter an obstacle elsewhere on its arm. Multi-Contact Feedback MPC accounts for these contacts explicitly, rather than assuming that interaction occurs only at a known end-effector location. It combines motion and force objectives with limits on unwanted contact forces.

DLR-KUKA robot performing end-effector force control while a person contacts another link — End-effector motion and force control continue while the robot responds to an additional human contact.

An Explicit Contact Feedback Loop

MCP-EP estimates contact locations and forces from joint torque and base force/torque measurements. The controller incorporates this information into a spring contact model and predicts its effect on the full seven-degree-of-freedom robot dynamics. Other estimators can also supply the feedback if they provide multi-contact information in real time.

Control architecture connecting MCP-EP, contact-feedback MPC, and the torque-controlled robot — Contact estimation closes the feedback loop around the predictive controller.

Contact Modeling and Optimization

The spring model predicts force along the estimated contact direction while leaving tangential motion unconstrained. The optimization combines end-effector pose tracking, contact-force regulation, and a force barrier that penalizes approaching or exceeding the permitted force.

Estimated contact frame, spring stiffness, and predicted contact force — The spring contact model connects estimated contact locations to predicted interaction forces.

Box-FDDP solves the optimal control problem with analytic derivatives and a 125 ms prediction horizon. Reported update rates are 6.8 kHz without contact, 1.9 kHz with one contact, and 1.8 kHz with two contacts. The implementation retains the full robot dynamics without the simplifying assumptions discussed in the source paper.

Experiments on the DLR-KUKA LWR IV+

Limiting an Unexpected Contact

When an unknown obstacle prevents the arm from reaching its target, MPC without contact feedback produces a peak force of 65.87 N. With contact feedback and a 15 N force limit, the peak is 14.74 N.

Comparison with and without contact feedback near an unknown obstacle.

Hybrid Motion–Force Control

The robot tracks a circular end-effector trajectory while regulating contact force to 20 N. An additional human contact on the third link is limited to 15 N. The table reports tracking performance during the three phases of this experiment.

Four stages from initial free motion to hybrid control and an additional human contact — Scenario 2: approaching the environment, establishing contact, and responding to an additional contact.

Scenario 2 tracking RMSE.
Phase	Position	Force
Hybrid motion–force control	0.86 cm	0.58 N
Additional human contact	1.14 cm	0.89 N
Constant desired pose	0.20 cm	0.10 N

Hybrid motion–force control with an additional human contact.

Supplementary demonstration: transferring an object of unknown weight while limiting a second contact force.

Representative Publications

Online Multi-Contact Feedback Model Predictive Control for Interactive Robotic Tasks

Seo Wook Han, Maged Iskandar, Jinoh Lee, and Min Jun Kim

IEEE International Conference on Robotics and Automation (ICRA), 2024

Paper

Bib

@inproceedings{han2024online,
  title={Online Multi-Contact Feedback Model Predictive Control for Interactive Robotic Tasks},
  author={Han, Seo Wook and Iskandar, Maged and Lee, Jinoh and Kim, Min Jun},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  pages={11556--11562},
  year={2024}
}

Youtube Project

Figures, demonstrations, and research details: author's project page.