A line drawn through a circuit can be smooth, elegant and impossible to drive. Autonomous racing planning begins when geometry acquires time, vehicle dynamics and a response to uncertainty.
The planner receives an estimated state rather than a perfect map of reality. It must return something the controller can interpret: a sequence of positions, headings, speeds or other agreed reference quantities. That sequence must remain coherent even when a new observation forces a different decision a moment later.
Separate the route from the motion
At a high level, a route identifies which part of the circuit to follow. A path describes a geometric curve through that route. A trajectory adds timing. These distinctions matter because the same path may be feasible at one speed and infeasible at another.
Curvature links speed to lateral acceleration. On an idealised planar path, the familiar relationship is lateral acceleration equals speed squared multiplied by curvature. This is a useful first check, not a complete tyre model. Load transfer, aerodynamic forces, surface variation and combined braking change the available margin.
A speed profile must also respect what happens before and after a corner. If the next bend requires a lower entry speed, braking must begin early enough. Looking only at the local curvature can produce a sequence of individually plausible targets that cannot be connected by the vehicle.
Spend a limited force budget
Tyres produce longitudinal and lateral forces through the same contact patches. Heavy braking and aggressive turning therefore compete for capacity. A friction circle or ellipse is a convenient approximation for reasoning about that competition, although actual behaviour depends on the tyre and operating conditions.
The planner should use limits that match the controller’s model closely enough to avoid contradictory expectations. A trajectory that assumes instantaneous steering changes cannot be rescued by a controller connected to a finite-rate steering actuator. Similarly, an optimistic deceleration limit can move a braking point beyond recoverability.
Conservative margins are not simply a permanent reduction in ambition. They can reflect the confidence of the current estimate. An uncertain grip estimate or a partially observed boundary can justify more clearance and a lower speed until additional evidence becomes available.
Planning contract
A candidate motion is useful only if it arrives on time and remains meaningful when its assumptions weaken.
Optimise repeatedly, not once
Model predictive control, often shortened to MPC, illustrates a receding-horizon approach: optimise a finite future, apply part of the result and solve again after receiving new state information. Planning and control may use related methods, but their responsibilities and timescales should still be explicit.
The objective can include progress, tracking error, smoothness and distance from constraints. These terms do not automatically become comparable because they appear in the same equation. Units, scaling and weights influence the behaviour. A design review should ask what happens when a soft penalty conflicts with a hard safety constraint.
Optimisation also consumes time. A nominally excellent result that arrives after its deadline may describe a world that has already changed. The architecture needs a policy for solver failure, infeasibility and stale outputs. Reusing a previous plan is acceptable only within a defined validity period and after relevant checks.
Leave room for other agents
An opponent trajectory is a prediction, not a reservation. Another vehicle may brake unexpectedly, change its line or respond to the autonomous car’s own motion. Planning around a single deterministic forecast can conceal these interactions.
Multiple hypotheses or bounded occupancy regions can represent alternatives. Their usefulness depends on how those bounds were constructed and what evidence supports them. Extremely broad uncertainty may make progress impossible; extremely narrow uncertainty may make unsafe progress look efficient. Neither problem is solved by presenting a prediction with more decimal places.
The Indy Autonomous Challenge and its Dallara AV-21 platform provide a real context for high-speed autonomous competition. They do not establish a universal planning recipe. Vehicle configuration, event rules and operating conditions must be considered separately when interpreting public demonstrations.
Keep an exit from the decision
A planner should ask whether the chosen motion preserves a meaningful fallback. A locally collision-free path may still lead into a state from which the car cannot stop within the known free space. Terminal constraints, reachable-set reasoning or explicit stopping corridors can help express that concern.
Autoware offers an open-source autonomous-driving ecosystem worth studying for interfaces and architectural patterns, while its road-oriented components should not be assumed ready for racing. Transfer the questions before transferring the code: what is the model, what are its limits and which component rejects an invalid command?
The final product of planning is therefore more than a preferred line. It is a time-bounded proposal, supported by assumptions and paired with a response when those assumptions fail. The controller now has to turn that proposal into forces without losing the distinction between requested and achievable motion.