Aim
Compare explicit constraints and sampling on a shared vertical-flight model.
Time: about 120 minutes. Start with the linked lectures and the course notation contract. The reference program is complete and runnable; your work is to explain it, modify it, and test the results.
Run from this repository
uv sync --locked
uv run --locked python practices/p04-predictive-control/run.py
The program writes configuration, MuJoCo version, and metrics to outputs/p04-predictive-control/metrics.json. Experiments with a time history also write trajectory.csv. These are generated results and are not source material.
Use a separate configuration and output directory for each variation:
uv run --locked python practices/p04-predictive-control/run.py --config practices/p04-predictive-control/config.yml --output outputs/p04-predictive-control/baseline
Keep the original configuration as the baseline. Record all changes in your report; avoid changing several physical and numerical parameters at once.
Tasks
- Run clipped LQR, constrained MPC, and the MPPI-inspired weighted-shooting baseline.
- Compare RMS height error, state-limit violations, solve failures, and compute time.
- Sweep horizon, temperature, and sample count using separate YAML files.
- Initialize a case outside the recoverable envelope and explain feasibility rather than hiding the failed solve.
Expected checks
The model is a one-axis double integrator around hover, not a complete quadrotor. MPC checks predicted height constraints. Weighted shooting penalizes violations but does not guarantee feasibility. Timing values depend on the machine.
Submission
Submit your YAML configuration, a short explanation of the equations and assumptions, labeled plots or numerical comparisons with units, and a paragraph explaining a failure or limitation. Include the exact run command. The weekly practice rubric awards 40% for correctness and checks, 30% for the controlled comparison, 20% for interpretation, and 10% for reproducibility.
Extension
Use Practice 3's full vehicle in TH1. A general-purpose SLSQP solve teaches formulation; a dedicated QP solver is a subsequent engineering improvement.