humanoidsdata.com

Search

Search companies, datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Simulation & transfer

System identification

System identification is the process of estimating a dynamical model from measured inputs and outputs. In robotics it can fit parameters of a known mechanical model or learn a model structure that predicts how a robot or environment responds to actions.

Also known as: robot system identification, dynamical system identification

Updated

Learn the response, not just the current state

A state estimator asks where the robot is and how it is moving. System identification asks which model explains its response to inputs. The problems can interact, but a joint-angle estimate is not itself an identified actuator or rigid-body model.

Tedrake's system-identification notes distinguish fitting parameters when the mechanical structure is known from learning more general input-output or state-space models. On a humanoid, structured identification might estimate combinations of inertial parameters, damping or joint friction. It is not restricted to fitting a simulator, although identified parameters can improve one.

More recordings do not make every parameter identifiable

Different parameter values can produce indistinguishable measured behaviour. The MIT notes explain that robot dynamics often reveal only identifiable combinations of parameters, not every physical parameter independently. They also discuss experiment design: the recorded motions must excite the effects being estimated, rather than repeatedly collecting the same uninformative trajectory.

For example, fitting a joint model only from near-stationary recordings is not a convincing validation of its predictions during rapid motion. This is a data-selection implication, not a claim that aggressive excitation is always appropriate. Identification experiments still need hardware limits, contact conditions and safe operating bounds.

Sensor calibration and timestamp alignment matter because biased or misaligned measurements can be absorbed into the fitted dynamics. Preserve the original measurements and document any filtering, derivative estimation and conversion from motor signals to joint effort.

Validate rollouts as well as one-step predictions

Tedrake distinguishes equation error from simulation error: a model can fit one-step transitions well yet accumulate large errors when its own predictions are fed forward. Evaluation should therefore include held-out trajectories and prediction horizons relevant to the intended controller, not just training loss.

System identification also differs from domain randomisation. Identification fits measured behaviour; randomisation trains or tests across a distribution of conditions. The methods can be combined by centring variation around an identified model while retaining uncertainty that the data did not resolve.

A useful identification dataset records commands, measured response, timing, units, operating conditions and the train/validation split. Publish the model's intended regime alongside its parameters: a fit for an unloaded joint is not evidence that the same model captures every payload, temperature or contact condition.

Sources