Adaptive Dynamics Orchestration boosts off-road robot navigation accuracy
New framework dynamically selects the best model for terrain, reducing errors by 40%
Researchers propose Adaptive Dynamics Orchestration (ADO), a framework that dynamically selects the most appropriate dynamics model for Model Predictive Control (MPC) in autonomous off-road navigation. ADO maintains a library of models with varying accuracy-efficiency trade-offs and uses online counterfactual rollouts to refine performance estimates. Real-world experiments on an off-road ground robot show ADO significantly reduces modeling error compared to fixed low-latency baselines while approaching the accuracy of the highest-fidelity model without its computational cost.
- ADO maintains a library of dynamics models with diverse accuracy-efficiency profiles, selecting the best one for the current terrain in real time
- Online counterfactual rollouts replay executed control actions across the model library to continuously refine performance estimates
- Real-world experiments on an off-road ground robot show ADO reduces modeling error compared to fixed low-latency baselines while approaching high-fidelity accuracy
Why It Matters
Enables safer, more reliable autonomous navigation in challenging outdoor environments without sacrificing real-time performance.