[Paper Review] Learning Compliance Adaptation in Contact-Rich Manipulation

Jianfeng Gao, You Zhou+1 more|arXiv (Cornell University)|May 1, 2020

Robot Manipulation and Learning10 references4 citations

TL;DR

This paper proposes a bi-directional GRU-based predictive model for force profiles in contact-rich manipulation, enabling real-time anomaly detection and adaptive impedance control. The system achieves high-force tracking accuracy during normal operation and rapidly adapts compliance to external disturbances, with abnormal events detected in under 70ms.

ABSTRACT

Compliant robot behavior is crucial for the realization of contact-rich manipulation tasks. In such tasks, it is important to ensure a high stiffness and force tracking accuracy during normal task execution as well as rapid adaptation and complaint behavior to react to abnormal situations and changes. In this paper, we propose a novel approach for learning predictive models of force profiles required for contact-rich tasks. Such models allow detecting unexpected situations and facilitates better adaptive control. The approach combines an anomaly detection based on Bidirectional Gated Recurrent Units (Bi-GRU) and an adaptive force/impedance controller. We evaluated the approach in simulated and real world experiments on a humanoid robot.The results show that the approach allow simultaneous high tracking accuracy of desired motions and force profile as well as the adaptation to force perturbations due to physical human interaction.

Motivation & Objective

• To enable robots to maintain high stiffness and precise force tracking during normal contact-rich manipulation tasks.
• To detect abnormal events—such as collisions, human interruptions, or loss of contact—using predictive force modeling.
• To dynamically adapt compliance in response to detected anomalies while ensuring safe recovery.
• To enable real-time, low-latency adaptation in both simulated and real-world humanoid robot experiments.
• To combine predictive force modeling with adaptive impedance control for robust, safe human-robot interaction.

Proposed method

• A bidirectional gated recurrent unit (Bi-GRU) network is trained on force and motion data from normal task executions to learn predictive models of expected force profiles.
• The predictive model outputs a mean force estimate and a 3-sigma confidence interval to quantify uncertainty.
• Anomalies are detected when the actual force measurement falls outside the predicted confidence interval, triggering an abnormality score.
• A threshold-based abnormality detection mechanism (e.g., −1000) triggers compliance adaptation within 70ms.
• An adaptive impedance controller reduces stiffness to 2% of normal value within 300ms upon anomaly detection, enabling compliant behavior.
• After abnormal events, the system gradually recovers control parameters to normal values to prevent instability, ensuring safe resumption of high-stiffness operation.

Experimental results

Research questions

• RQ1 Can a predictive force model based on Bi-GRU effectively detect deviations from expected force profiles during contact-rich manipulation?
• RQ2 How quickly can the system detect and respond to abnormal events such as collisions or human interference?
• RQ3 Can adaptive compliance be achieved without sacrificing force tracking accuracy during normal task execution?
• RQ4 How does the system ensure stable and safe recovery after abnormal events without inducing instability?
• RQ5 Can the proposed method generalize across different surface geometries and dynamic interaction scenarios?

Key findings

• The system detected abnormal events—including collisions, interruptions, and human interactions—in less than 70ms after occurrence.
• During normal operation on flat, sloped, and curved surfaces, the root mean squared error (RMSE) of force regulation was 1.047N, indicating high tracking accuracy.
• The predicted force profile closely matched actual measurements, with confidence intervals covering 99.7% of the data under the 3-sigma rule.
• Stiffness was reduced to 2% of its normal value within 300ms of anomaly detection, enabling compliant behavior during disturbances.
• Recovery of control parameters was performed gradually, preventing jiggling and instability, especially during successive abnormal events.
• The method was successfully validated on the ARMAR-6 humanoid robot in both simulation and real-world experiments across multiple surface types.