A simulated two-link robot arm must reach three targets in order while avoiding an obstacle. Its physics are mostly known: gravity, inertia, and how the two links pull on each other. What the equations miss is an unknown payload, sticky friction, and motors that deliver less torque than commanded.
A planner (model-predictive control) picks each move by simulating it with a model, so every model error becomes a planning error. With only a little measured data, the study compares three ways to fix the model: re-fit its physical parameters, replace it with a neural network, or let a network learn only what the model gets wrong (the residual). Each is judged by how often the arm completes its task, not by how well it predicts the next step.
The simulator, the models, the planner, and the analysis already run end to end on a small test setup. The task itself is still being calibrated, so there are no results yet. The analysis rules are fixed before the run, and a null or opposite result will be reported the same way.
Context
Primary references
- Zeng et al. (2020): TossingBot (learning residual physics on top of an analytical model)
- Ajay et al. (2018): Augmenting physical simulators with stochastic neural networks
- Deisenroth & Rasmussen (2011): PILCO: data-efficient model-based policy search
- Rawlings, Mayne & Diehl: Model Predictive Control: Theory, Computation, and Design
Novelty scope: Learning a correction on top of a physics model is an established idea. This project is a carefully controlled comparison, not a new algorithm; its planned contribution is the rigor of that comparison.