Feedback compensation through state rollouts and cost queries
1Southeast University 2The Hong Kong University of Science and Technology (Guangzhou) 3Shanghai University of Electric Power
Project video
A visual introduction to disturbance observation through rollout and cost-query interfaces.
Overview
SDOB extends disturbance observation to analytical models, simulators, and learned predictors accessed through state rollouts or cost queries.
Robotic controllers increasingly rely on analytical models, simulators, cost-query interfaces, and learned world models. However, physical deployment can deviate from nominal assumptions, and additional disturbances may arise even when the model itself is accurate. In control systems, disturbance observers (DOB) are widely used to estimate such unmeasured effects from nominal models and measured feedback. Classical DOB formulations are generally built around explicit plant models.
This paper develops the sampling-based disturbance observer (SDOB), extending the DOB principle to a broader range of models, including simulators and learned world models, through state-rollout or cost-query interfaces. SDOB separates two observable channels: state-effect disturbances, for which the feedback state differs from its prediction, and cost disturbances, for which the same query state receives different costs as the perceived environment changes. Diverse simulation and real-robot experiments across traditional and learned models demonstrate the effectiveness of SDOB in compensating for model mismatch and improving control performance.
Contributions
SDOB targets locally persistent, task-relevant mismatch through the interfaces used for prediction and planning.
We formulate disturbance compensation at two predictive-control interfaces. The state channel represents feedback mismatch through a task-defined correction model, while the cost channel represents local motion of an environment-dependent cost field.
A forward-evaluation implementation estimates these corrections from successive feedback and incorporates them into future state propagation and cost evaluation. Given the required correction and query interfaces, SDOB does not require analytic model inversion or retraining of the nominal model.
Five simulation studies and three representative real-robot demonstrations evaluate complete SDOB-equipped controller configurations across analytical and learned models. Source code, configurations, and simulation scripts are released to support reproducibility.
Two channels
The state channel fits feedback transitions within a task-defined disturbance model. The cost channel estimates local motion of an environment-dependent cost field.
State channel
The feedback state differs from its nominal prediction. SDOB samples corrections within a task-defined disturbance model, evaluates them through forward prediction, and scores how well they explain the completed transition.
Cost channel
The same query state receives a different cost after the perceived environment changes. SDOB matches local cost patterns across successive feedback updates to estimate cost-field motion for future rollout evaluation.
State-effect compensation
Candidate disturbances are propagated through the forward model and scored against the completed feedback transition. Regularization and physical admissibility suppress implausible explanations, and the resulting estimate corrects the next state rollout.
Cost transport
Local query matches estimate how the observed cost pattern moved. Interpolation makes that estimate available at arbitrary rollout states, allowing the controller to evaluate where the high-cost region is expected to move.
Control integration
Compare completed model predictions and cost queries with feedback.
Fit task-defined state corrections and local cost-field motion using forward evaluations.
Correct subsequent state propagation and the costs used to score candidate rollouts.
Simulation
The studies compare complete controller configurations across analytical and learned models using 20 paired trials or seeds. They assess task performance under the specified disturbances.
Analytical model · State effect
SDOB estimates a matched torque bias from the completed state transition and compensates the next MPPI rollout.
Robot cost model · Both channels
A constant input bias changes vehicle motion while moving obstacles change collision costs at the same query states.
Latent world model · State effect
The frozen LeWorldModel predicts latent rollouts from images. SDOB estimates an equivalent action bias from latent endpoint mismatch. Evaluation uses 20 tasks that disturbance-free MPPI can already solve.
MMBench2 world model · Cost channel
The world model supplies planning scores, while reconstructed images supply obstacle-distance cost maps. SDOB estimates local map motion and compensates future costs with asteroid speed increased to 1.5× nominal.
MMBench2 world model · Both channels
With an unknown horizontal input bias and deck speed increased to 1.6× nominal, SDOB corrects state predictions and shifts future landing targets. All four methods reach the deck; the main benefit is improved touchdown alignment.
Metrics report MPPI+SDOB unless a baseline is named. These comparisons evaluate complete controller configurations; they do not separately isolate each channel or correction prior.
Hardware
Vehicle and dual-arm obstacle avoidance demonstrate feedback compensation in closed loop. PointWorld box lifting uses initial feedback calibration followed by execution of a single plan.
Real-car navigation
MPPI+SDOB completed the representative route in 9.29 s, the shortest time among the shown trials.
Dual-arm manipulation
The obstacle pose updates the collision model, but its velocity is not supplied. SDOB estimates motion from changing collision-cost queries and begins avoidance earlier.
PointWorld · Initial feedback calibration
An initial 50 mm lift provides feedback to fit a task-specific correction to PointWorld predictions. MPPI then plans a 300 mm lift using the calibrated predictor. The selected action sequence executes without replanning or observer updates.