Rapid Mismatch Estimation

Abstract

With robots increasingly operating in human-centric environments, ensuring soft and safe physical interactions, whether with humans, surroundings, or other machines, is essential. While compliant hardware can facilitate such interactions, this work focuses on impedance controllers that allow torque-controlled robots to safely and passively respond to contact while accurately executing tasks. From inverse dynamics to quadratic programming based controllers, the effectiveness of these methods relies on accurate dynamics models of the robot and the object it manipulates. Any model mismatch results in task failures and unsafe behaviors. Thus, we introduce Rapid Mismatch Estimation (RME), an adaptive, controller-agnostic, probabilistic framework that estimates end-effector dynamics mismatches online, without relying on external force-torque sensors. From the robot's proprioceptive feedback, a Neural Network Model Mismatch Estimator generates a prior for a Variational Inference solver, which rapidly converges to the unknown parameters while quantifying uncertainty. With a real 7-DoF manipulator driven by a state-of-the-art passive impedance controller, RME adapts to sudden changes in mass and center of mass at the end-effector in ~400 ms, in static and dynamic settings. We demonstrate RME in a collaborative scenario where a human attaches an unknown basket to the robot's end-effector and dynamically adds/removes heavy items, showcasing fast and safe adaptation to changing dynamics during physical interaction without any external sensory system.

Static Experiments

To evaluate RME, we extensively tested the model with multiple static experiments, where the physical manipulator, subject to sudden changes in the dynamics model resulting from adding unknown mass to the end-effector, aimed to maintain the target equilibrium position and orientation. We compare the response of nominal Constrained Passive Interaction Controller (CPIC) to the same controller with RME gravity compensation augmentation. RME provided rapid and accurate estimation of the attached external load, allowing the robot to quickly converge to the equilibrium position from before applying the mismatch.

Tracking a Stable Limit Cycle

To test the model in the dynamic scenario, we designed a Dynamical System with a stable limit cycle in the y-z plane, where the manipulator aims to achieve target velocities from the motion planner. As shown, RME rapidly estimates mismatch parameters, allowing the robot to converge to the desired trajectory. Without adaptation, the manipulator reaches the spurious attractor, where the task force from the impedance controller is balanced by the gravitational force acting on the mass mismatch, preventing it from continuing on the given task.

Sequential Adaptation with Human-Robot Interactions

In this experiment, we evaluate the model's capability to perform continuous adaptation while maintaining passivity with respect to human-generated perturbations. We attach the unknown basket to the end effector, sequentially add and remove two unknown objects, and finally, remove the basket, perturbing the robot between these actions. RME provided an accurate load estimation, allowing the robot to maintain target position and orientation, while being passive to human-generated forces (even when the model confused human perturbations with model mismatch, it allowed perturbation, and instantaneously corrected itself after being released). Without RME-based adaptation, robot is pulled down by the gravitational force of external objects, which activates joint limit constraint of the Constrained Passive Interaction Controller.

Pick and Place with Human-Robot Interactions

In this experiment, we evaluate framework performance in the Pick and Place task with a passive velocity-based inverse kinematics (PVIK) controller, utilizing the qb SoftHand2 Research end effector, and the OptiTrack motion capture system. In the first test, the robot aims to pick an unknown object (700 grams), and place it on top of the box, maintaining passivity with respect to human-generated perturbations. During the process, we add an additional 500 grams to the basket. RME allows the robot to track the desired trajectory by rapidly adapting to unknown dynamics, while a controller without RME fails to pick up a heavy object. In the second test, the robot aims to track and intercept the basket (1200 grams) from a human and place it on top of the box, which is possible with the use of RME. In both experiments, the controller with RME augmentation preserves stability and passivity of the closed-loop system.

Acknowledgments

We thank our anonymous reviewers, who provided thorough and fair feedback that improved the quality of our paper. This work was supported by the National Science Foundation (NSF) Foundational Research in Robotics (FRR) program under NSF CAREER Award Grant No. FRR-2443721.

@inproceedings{jaszczuk2025rme, author = {Jaszczuk, Mateusz and Figueroa, Nadia}, title = {Rapid Mismatch Estimation via Neural Network Informed Variational Inference}, booktitle = {9th Conference on Robot Learning (CoRL)}, year = {2025} }

Rapid Mismatch Estimation via Neural Network Informed Variational Inference