When a robotic system operates in the real world, it rarely possesses a perfect model of its environment or even its own physical state. A drone might pick up an unexpected payload; an exoskeleton might experience varying joint friction; a mobile manipulator might interact with surfaces of unknown stiffness.
Traditional control strategies handle these discrepancies in one of two ways. Robust control designs a single, conservative controller that guarantees safety across a worst-case range of parameters, but sacrifices performance. Offline system identification builds highly accurate models beforehand, but fails the moment conditions change during operation.
Online parameter estimation offers a elegant middle ground. Instead of treating model mismatch as unmodeled noise or a static error, the system treats physical properties—such as mass, inertia, friction, and stiffness—as dynamic states to be inferred in real time. By continuously updating its internal physical model based on observed behavior, the controller can dynamically adapt its strategy, maintaining high performance even in highly volatile environments.
The Estimation-Control Loop
The core intuition behind online parameter estimation is that a system's tracking errors and deviations from expected trajectories contain structured information about its true physical properties. If a robotic arm accelerates slower than predicted given a specific motor torque, the system can infer that either the mass or the joint friction is higher than initially assumed.
This insight forms a closed loop consisting of an estimator and a functional controller. The estimator observes short histories of states and actions, isolates the discrepancies between predicted and actual outcomes, and updates its parameter estimates. The controller then uses these updated parameters to adjust its control laws.
[ Control Inputs ] ---> ( Physical System ) ---> [ Observed States ]
^ |
| v
( Updated Model ) <--- [ Parameter Estimator ] <--- [ Prediction Error ]
This architecture traces back to foundational work in neural and parallel processing. Gomi et al. in Recognition of Manipulated Objects by Motor Learning demonstrated how an estimation network could predict continuous physical parameters like mass or viscosity to modulate a functional controller, allowing a system to generalize across a continuous range of environmental variations. To make this viable for high-performance mechanical systems, Morris (Identification of Robot Dynamics: A Parallel Processing Approach) showed that distributing the computational load of Lagrangian mathematical models across parallel processors could achieve the high-speed, online estimation of link inertial parameters required during active motion.
In modern applications, this estimation often relies on recursive filtering techniques. For example, Grandia (Model Predictive Robot-Environment Interaction Control for Mobile Manipulation Tasks) utilizes a recursive filter to identify environmental properties like mass and stiffness during physical contact, feeding these values directly back into a predictive model. Similarly, Byun et al. (Stable Contact Guaranteeing Motion/Force Control for an Aerial Manipulator on an Arbitrarily Tilted Surface) employ a recursive least-squares estimator to identify surface stiffness and damping parameters in real time, allowing an aerial manipulator to maintain stable contact against unknown materials.
Enhancing Predictive and Learning-Based Frameworks
Online parameter estimation acts as a powerful plug-in for advanced control frameworks like Model Predictive Control (MPC) and Reinforcement Learning (RL), both of which are highly sensitive to model accuracy.
Model Predictive Control
MPC relies on rolling-horizon optimization to plan future actions based on a forward model. If the forward model is wrong, the plan degrades. The authors of Adaptive CLF-MPC With Application to Quadrupedal Robots address this by using an online regressor to estimate a quadrupedal robot's mass and center of gravity, allowing it to carry heavy, off-center loads without tipping over.
For safety-critical systems, Yuan et al. (Safety-based Speed Control of a Wheelchair using Robust Adaptive Model Predictive Control) integrate online parameter estimation into a robust MPC framework for electric wheelchairs, automatically adjusting to significant load variations (such as different user weights) to prevent dangerous accelerations. Taking adaptation a step further, Oshin et al. (Differentiable Robust Model Predictive Control) apply the Implicit Function Theorem to tube-based MPC, updating the controller's robustness parameters online via gradient descent to dynamically balance conservative safety margins against aggressive tracking performance.
Reinforcement Learning
In learning-based control, online estimation helps bridge the "reality gap" between simulation and the real world. Yu et al. (Preparing for the Unknown: Learning a Universal Policy with Online System Identification) introduced the UP-OSI framework, which pairs a universal policy with an online system identification module. The module infers environmental parameters (like friction or mass) on the fly from short state-action histories, allowing the policy to generalize to physical setups it never encountered during training.
Alternatively, instead of running a separate identification module, the physical parameters can be embedded directly into the control state. Hu et al. (Automatic Control Optimization for Large-Load Plant-Protection Quadrotor) utilize an "extended-state" RL approach that incorporates changing moments of inertia—such as a diminishing liquid payload during flight—directly into the agent's state, outperforming traditional static controllers.
Ensuring Stability Amidst Adaptation
A major risk of updating a controller's parameters online is instability. If noisy sensor data causes the estimator to fluctuate wildly, the controller might respond with erratic inputs, creating a catastrophic feedback loop.
To prevent this, researchers rely on indirect adaptive control, where update laws are mathematically constrained using Lyapunov stability theory. Instead of directly tuning controller gains based on performance, the system updates the underlying physical parameters of the model in a way that guarantees the tracking error will asymptotically converge to zero.
Yang et al. (Indirect Adaptive Fuzzy Observer Design) leverage this approach by deriving parameter update laws via Lyapunov stability theory, ensuring that an internal model self-corrects safely even when the underlying nonlinear system dynamics are completely unknown. This theoretical guarantee enables robust real-world deployment across complex mechanical architectures:
- Exoskeletons: Zhang et al. (A hierarchical Lyapunov-based cascade adaptive control scheme for lower-limb exoskeleton) integrate adaptive estimation algorithms within a back-stepping framework to identify unknown mass, inertia, and friction in real time, maintaining high tracking accuracy without requiring a perfect prior hardware model.
- Multi-Robot Collaboration: Yan (Decentralized Ability-Aware Adaptive Control for Multi-robot Collaborative Manipulation) applies Lyapunov-stable adaptation laws to a decentralized team of diverse robots. The individual agents estimate unknown object properties (like mass and friction) online, ensuring system-wide stability and coordination without requiring constant inter-robot communication.
Generalization Beyond Rigid Bodies
The utility of online parameter estimation extends beyond rigid mass and inertia matrices. It is increasingly applied to highly non-linear, deformable materials where analytical physics models are impossible to construct perfectly.
For instance, tracking and manipulating fabric presents severe challenges due to complex internal friction and stiffness variations. Held et al. (Self-supervised Cloth Reconstruction via Action-conditioned Cloth Tracking) address this by performing a localized online search for material properties during the tracking process. By dynamically calibrating stiffness and friction parameters to match observed cloth movements in real time, their physics-based tracker can generalize across completely different fabric types on the fly.
Whether dealing with a shifting liquid payload, a changing human user, or the intricate folds of a garment, online parameter estimation transforms the control problem. By turning model uncertainty into an observable, time-varying state, it allows robotic systems to safely and precisely navigate the unpredictable dynamics of the real world.
Key Foundational Work
Identification of Robot Dynamics: A Parallel Processing Approach (1989) — contributed 1 concept elements to this direction.
Recognition of Manipulated Objects by Motor Learning (1991) — contributed 1 concept elements to this direction.