Source-linked AI summary
Variable impedance control and learning -- A review
Fares J. Abu-Dakka, Matteo Saveriano
TL;DR
Robots interacting with changing, uncertain environments need safe ways to regulate contact forces and adapt their behavior online. The survey synthesizes variable impedance control and learning approaches, proposes a taxonomy, and concludes that a combined framework could unite controller robustness with learning flexibility, although theoretical and practical difficulties remain.
Problem
Changing interaction conditions require robots to adapt impedance online, while existing approaches differ in their control, learning, and safety properties.
Method
The survey reviews variable impedance control, variable impedance learning, and their combination, organizing the field through a taxonomy based on variability, learning, and control.
Results
The review identifies control, learning, and combined VILC approaches, with varying advantages and disadvantages, and presents VILC as the route toward an omni-comprehensive variable impedance framework.
Takeaways & Limitations
The envisioned framework should combine the accuracy and robustness of properly designed controllers with the flexibility and generalization of learning approaches.
Takeaways & Limitations
Reinforcement-learning approaches can be highly data inefficient, and the reviewed approaches do not guarantee safe exploration.
Abstract
from arXiv · showhide
Robots that physically interact with their surroundings, in order to accomplish some tasks or assist humans in their activities, require to exploit contact forces in a safe and proficient manner. Impedance control is considered as a prominent approach in robotics to avoid large impact forces while operating in unstructured environments. In such environments, the conditions under which the interaction occurs may significantly vary during the task execution. This demands robots to be endowed with on-line adaptation capabilities to cope with sudden and unexpected changes in the environment. In this context, variable impedance control arises as a powerful tool to modulate the robot's behavior in response to variations in its surroundings. In this survey, we present the state-of-the-art of approaches devoted to variable impedance control from control and learning perspectives (separately and jointly). Moreover, we propose a new taxonomy for mechanical impedance based on variability, learning, and control. The objective of this survey is to put together the concepts and efforts that have been done so far in this field, and to describe advantages and disadvantages of each approach. The survey concludes with open issues in the field and an envisioned framework that may potentially solve them.
1 INTRODUCTION
Robots increasingly interact with dynamic, uncertain, and human-inhabited environments, motivating impedance-based methods that regulate motion and compliance in response to forces. The survey organizes variable impedance control and learning approaches and focuses on their adaptation to changing interaction conditions.
- Motivation: Unstructured environments require interaction methods that accommodate dynamic, uncertain conditions and possible human presence.Impedance control is presented as an advanced interaction methodology for such settings.
- Impedance control: Impedance control modulates robot motion or compliance according to force perceptions, helping avoid large impact forces under position uncertainty.It is distinguished from compliance control, which is described as a subset of impedance control.
- Impedance control: An impedance controller resembles a virtual spring–damper between the environment and end-effector, using mass, stiffness, damping, position, orientation, velocity, force, and torque quantities.The standard interaction model includes translational and rotational impedance components.
- Implementation: Impedance control can be implemented in Cartesian or joint space, while admittance control maps sensed external forces into desired velocities when low-level torque or current access is unavailable.Admittance control is described as a practical dual strategy for position- or velocity-controlled robots equipped with additional force sensing.
- Variable impedance: Tasks such as turning valves or opening doors require impedance variation because different masses and friction forces demand different control forces.Sensed forces provide relevant information about the forces needed for these manipulation tasks.
- Survey scope: The survey reviews variable impedance control from control and learning perspectives and proposes a taxonomy organized around mechanical impedance, variability, learning, and control.The taxonomy separates constant and variable impedance gains and includes control, stability, and learning approaches.
2 VARIABLE IMPEDANCE CONTROL (VIC)
Variable impedance control adjusts impedance parameters during task execution to handle changing interaction demands and support safety, flexibility, and efficiency. The reviewed approaches include time-varying control laws, application-specific adaptation, human-in-the-loop methods, and stability analyses based on Lyapunov tools or passivity.
- 2 VARIABLE IMPEDANCE CONTROL (VIC): Variable impedance control varies impedance during execution to address the unknown nature of everyday environments and task-dependent interaction forces.The survey describes VIC as a control-based approach for adjusting impedance parameters during tasks.
- 2 VARIABLE IMPEDANCE CONTROL (VIC): VIC uses time-varying impedance quantities, with the subscript t indicating that the relevant parameters change over time.A block scheme for the resulting controller is shown in Figure 3.
- 2 VARIABLE IMPEDANCE CONTROL (VIC): A representative update rule increases joint stiffness when trajectory tracking error is high, while a small baseline stiffness helps prevent unsafe interaction.The stiffness vector contains joint-wise values, and α is a positive gain.
- Applications: VIC has been applied to human–robot cooperation, rehabilitation, grasping, and manipulation, including peg-in-hole tasks with tolerance below 0.1 mm.In grasping, stiffness is varied using an estimated friction coefficient to help avoid object slippage.
- Stability: Because changing impedance complicates stability analysis with unknown environments, reviewed methods use Lyapunov-based analysis or passivity-based energy conditions.Passivity-based methods constrain dissipated and injected energy, while passivity-preservation control supports stable VIC in joint and Cartesian spaces.
- 2.2 VIC with human-in-the-loop: Human-in-the-loop VIC must address uncertainty while maintaining interaction stability and supporting the human, including through tele-impedance based on real-time EMG estimates.Tele-impedance transfers human impedance regulation skills to a robot interacting with an uncertain environment.
3 VARIABLE IMPEDANCE LEARNING (VIL)
Variable impedance learning learns impedance parameters from demonstrations and retrieves them online from current robot measurements. These methods typically learn stiffness and damping, while the underlying controller executes the resulting impedance behavior.
- 3 VARIABLE IMPEDANCE LEARNING (VIL): Robot learning has expanded from trajectory following to learning, reproducing, and adapting variable impedance parameters as skills.This development has been enabled by torque-controlled robots and learning algorithms for impedance gains.
- Learning formulation: VIL learns a parameterized nonlinear mapping from demonstrations containing trajectories and interaction data to variable impedance gains.The learned mapping is represented using parameters θ and is retrieved during execution.
- Online execution: At run time, VIL retrieves desired stiffness and damping from current position, velocity, and force measurements.The gains may be saturated or rate-limited to support desired closed-loop properties such as stable interaction.
- Controller integration: Training data are typically supplied by an expert, for example through kinesthetic teaching, and remain independent of the underlying controller.A VIC then converts the learned gains into the desired impedance behavior during execution.
- Scope: VIL approaches generally learn variable stiffness and damping matrices, whereas several VIC approaches also perform inertia shaping.The survey notes that the boundary between learning and controller design can become unclear in VILC approaches.
VIL via Imitation Learning
Imitation-learning approaches extend robot teaching from trajectories to variable impedance behavior, including stiffness and damping represented as structured SPD matrices. The surveyed methods range from probabilistic and demonstration-based estimation to direct manifold-aware learning of impedance gains.
- Learning variable impedance: Imitation learning has expanded from trajectory following to encoding variable impedance skills from demonstrations.Recent work also enables a novice robot to learn variable impedance behavior from an expert robot during online collaborative execution.
- Probabilistic demonstration learning: Gaussian mixture methods estimate attractors, stiffness matrices, or full stiffness profiles from position, force, visual, and haptic demonstrations.Approaches include time-driven GMM regression, GMR-based full stiffness estimation, and task-parameterized GMMs with weighted least squares.
- Geometric structure: Full stiffness and damping matrices are SPD, so learning algorithms must respect their geometric constraints rather than treating them as unconstrained vectors.Earlier approaches processed impedance matrices before and after learning, motivating direct learning without reparametrization.
- Direct SPD learning: Direct tensor-based learning of SPD force and stiffness data reproduced skills more accurately than Euclidean learning after Cholesky reparametrization.The comparison concerned tensor-based GMM/GMR versus traditional Euclidean-based GMM/GMR.
- Manifold-aware adaptation: Riemannian Dynamic Movement Primitives can operate on the SPD manifold and adapt impedance behavior to a new goal SPD point.This formulation follows the direct-manifold treatment of impedance data.
4 VARIABLE IMPEDANCE LEARNING CONTROL (VILC)
Variable impedance learning and control integrates learning with impedance-based robot interaction, often coupling parameter adaptation with the control structure. The reviewed approaches span imitation, iterative, and reinforcement learning, while highlighting trade-offs among flexibility, sample efficiency, transferability, stability, and safe exploration.
- VILC overview: VILC methods tightly integrate impedance learning with control, so the data-collection process and underlying controller can jointly determine learned behavior.Unlike VIL, VILC may require iterative updates or robot self-exploration and can also update the target trajectory.
- VILC via Imitation Learning: Imitation-based VILC uses demonstrations or multimodal human input to learn trajectory and stiffness parameters online, but some setups require complex calibration.Examples combine DMP updates with EMG, force, position, haptic, or muscular-activity measurements.
- VILC via Iterative Learning: Iterative learning control adapts control inputs across repetitions, but learning variable impedance requires replacing trajectory-error feedback with an error describing impedance mismatch.Conventional ILC with fixed target impedance may become unstable when the environment changes significantly.
- VILC via Reinforcement Learning: State-dependent impedance policies can improve sample efficiency and performance in contact-rich manipulation, while specialized parameterizations may reduce training time or enable transfer.The reviewed text reports increased sample efficiency and overall performance for specialized policies, and notes that some control policies transfer between simulation and real robots.
- Stability in VIC exploration: Safe exploration remains unresolved: stable policy parameterizations support stable exploration, but their constrained spaces may prevent finding an optimal policy.The survey suggests jointly updating policy parameters and the Lyapunov function to enlarge the safe set, while identifying this as requiring further investigation.
5 DISCUSSION
The survey reviews learning and control approaches for variable impedance controllers, identifies their complementary strengths and limitations, and proposes variable impedance learning and control as a route toward a comprehensive framework. Major unresolved issues include generalization, safety, theoretical guarantees, and the difficulty of combining accurate control with flexible learning.
- The review compares the main learning and control approaches for variable impedance controllers and summarizes their general advantages and disadvantages.
- An envisioned framework would combine controller accuracy and robustness with the flexibility and generalization of learning approaches.
- Stability and robustness guarantees are difficult for robots interacting with unstructured environments, while simplifying assumptions restrict the application scope of existing approaches.Passivity offers relatively general assumptions, but remains model-based and may still require simplifications for analytical solutions.
- Learning-based methods handle complex scenarios under minimal assumptions, but long training times and potentially unsafe behaviors remain robotics limitations.Safe and model-based reinforcement learning has begun addressing these issues, although results remain preliminary.
- Variable impedance learning and control is presented as a route toward an comprehensive variable impedance framework, but further theoretical and practical challenges remain.
- Imitation learning can acquire diverse impedance tasks from demonstrations, but failures may result from poor demonstrations, human–robot dynamics mismatch, or insufficient demonstrations for generalization.Combining imitation learning with trial-and-error, reward-driven learning is proposed to improve generalization to unseen task situations.
6 CONCLUDING REMARKS
The survey compares control- and learning-based approaches for varying robot impedance during task execution, highlighting complementary strengths and unresolved limitations. It identifies manifold and reinforcement learning as promising directions for learning effective variable impedance behaviors across industrial and service scenarios.
- Concluding remarks: Control and learning approaches each offer distinct advantages and disadvantages for adjusting impedance parameters during task execution.The survey categorizes traditional variable impedance control alongside variable impedance learning (VIL) and variable impedance learning control (VILC).
- Concluding remarks: Control approaches provide solid mathematical foundations, robustness, and efficiency but require substantial prior knowledge, whereas learning approaches often need less prior information but are data- and computationally inefficient.These limitations reduce the applicability of current variable impedance approaches in dynamic and unstructured environments.
- Concluding remarks: Manifold and reinforcement learning are identified as promising ways to overcome existing VILC limitations and learn effective variable impedance behaviors for industrial and service scenarios.The proposed potential scope includes both industrial and service applications.