---

name: mpc-optimization description: "Configures cuRobo trajectory optimization, rollout costs, and CUDA MPC loops for smooth, collision-aware robot control." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Optimization and MPC

Use this route for B-spline trajectory optimization, MPC action sequences, custom costs/rollouts, optimizer stages, or reactive control design. Read api-reference.md and workflows.md before constructing internal configs.

Core workflow

1. Choose TrajectoryOptimizerCfg.create for offline trajectory generation or ModelPredictiveControlCfg.create for receding-horizon control. Keep robot, optimizer, rollout, transition, metrics, and scene YAMLs from compatible families.
2. Configure DeviceCfg, collision/self-collision, tolerances, horizon/dt, seeds, and use_cuda_graph=True. Construct only after a CUDA probe.
3. For trajopt, call solve_pose, solve_cspace, or solve_state; inspect convergence, costs, collision metrics, and interpolated trajectory.
4. For MPC, update current state and goal/tool pose, then call optimize_next_action or optimize_action_sequence. Warm-start from the previous solution and use safe deceleration on failure.
5. Disable graphs only in a bounded debug/test pass. The included scripts/optimization_smoke.py checks configuration and optimizer construction without a long control loop.

For scene schemas use collision-scenes; for a composed planner use motion-planning.