Dynamic Parameter Identification in Robotic Systems

Dynamic parameter identification determines the physical constants that characterise a robot’s response to motion and external loads, encompassing quantities such as link masses, centres of mass, inertia tensors, joint friction coefficients and stiffness. Accurate estimation of these parameters underpins high-performance control, enabling precise trajectory tracking, energy-efficient motion planning and compliant interaction with uncertain environments. Traditional model-based approaches construct linear identification models in the parameters by applying excitation trajectories designed to sufficiently stimulate all dynamic modes; data are then processed via least-squares, instrumental variables or maximum likelihood estimators. Recent advances have enriched this classical framework through global optimisation algorithms, which incorporate physical consistency constraints to avoid unphysical parameter sets, and have separated parameter classes—such as stiffness versus inertial quantities—to mitigate coupling effects in elastic joints. Measurement noise, unmodelled flexibility and thermal drifting of friction remain key challenges, motivating sensorless and noise-robust schemes, as well as the integration of data-driven techniques. In particular, machine-learning methods have begun to compensate residual torque errors by learning unmodelled dynamics from experimental data, offering improved prediction accuracy across a variety of robot architectures. Together, these developments are shaping a more reliable and adaptable parameter identification landscape, with broad implications for industrial automation, robotic surgery, assistive devices and autonomous exploration.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent work has introduced hybrid schemes that combine classical inverse dynamics models with deep learning. In one study, a dual-stage identification process first estimates initial parameters via standard least-squares fitting, followed by a neural network that learns residual torque errors using long short-term memory and attention mechanisms; this approach achieved torque prediction errors below 6% on a six-degree-of-freedom manipulator. A broad survey and benchmarking toolkit has also been released, implementing seventeen established identification algorithms—from weighted least-squares and instrumental variables to recurrent neural networks and semi-definite programming methods—and systematically comparing their robustness, precision and computational efficiency on both simulated and real industrial robots. Additionally, global optimisation frameworks have been applied to enforce physical consistency constraints—such as positive-definite link inertia tensors and mass centre bounds—during parameter retrieval; validation experiments on collaborative platforms confirmed that constraint-aware optimisation yields accurate torque predictions across diverse motion profiles, highlighting its relevance for safe human–robot interaction and predictive maintenance.

Dynamic Parameter Identification in Robotic Systems publication trend

The graph below shows the total number of articles in dynamic parameter identification in robotic systems across all publications each year (not limited to Nature Index journals).

Technical terms

Dynamic parameter identification: The process of determining a robot’s mass, inertia, friction and stiffness characteristics by analysing motion and force data.

Inverse dynamics: A computation that estimates the joint torques or forces required to produce a specified motion, based on a dynamic model of the robot.

Excitation trajectory: A prescribed motion path designed to sufficiently stimulate all dynamic modes of a robot, ensuring reliable parameter estimation.

Inertia tensor: A matrix representation of how mass is distributed within a rigid body, influencing its resistance to rotational acceleration about different axes.

Global optimisation: A computational strategy that searches for the best set of parameters by minimising a cost function under physical and mathematical constraints, avoiding local minima.

Machine learning compensation: The use of data-driven algorithms, such as neural networks, to model and correct residual errors in physics-based dynamic models caused by unmodelled effects.

References

1. Deep Learning Aided Dynamic Parameter Identification of 6-DOF Robot Manipulators. IEEE Access (2020).
2. Dynamic Identification of the KUKA LBR iiwa Robot With Retrieval of Physical Parameters Using Global Optimization. IEEE Access (2020).
3. A dynamic parameter identification method of industrial robots considering joint elasticity. International Journal of Advanced Robotic Systems (2019).