Autonomous driving: simulation that does not match the car

The chain from planning to control must be validated in the loop and on the car

What happens. An autonomous highway function passes through perception, planning and control. The path the planner computes has to be one the controller can follow, and both have to work in simulation, on a rig with real hardware and on the vehicle. The links between the stages are where trouble sits.

What it costs. Rework when a planned path cannot be driven, test days used to find which stage is at fault, and slow release cycles.

Why the usual fix fails. Building each stage on its own makes each look fine. Adding test kilometres is slow and finds rare cases rarely. A comfort margin on the planner hides that the controller lags.

How a fixed-price remote diagnosis checks it. From recorded drives and the simulation set-up I look at where planned and driven paths separate and which stage causes it. The fixed-price remote diagnosis tells you whether trajectory optimisation that includes the vehicle limits can make the chain consistent, and what validation it needs.

Background: Autonomous mobile navigation

Customer vehicle dynamics are unknown, so simulation and road differ

What happens. An autonomy software company runs on vehicles built by its customers. Mass, tyres, steering, brakes and drive are known only roughly. Simulation with the nominal vehicle drives well; on the road the same software behaves differently.

What it costs. Surprises at each new vehicle, long on-site tuning, and simulation results that do not carry over to the customer.

Why the usual fix fails. Asking the customer for a full data sheet does not give the real dynamics. Tuning the software for each vehicle by hand is slow. A generic model is too far from the real one.

How a fixed-price remote diagnosis checks it. From a short set of manoeuvres recorded on the vehicle I look at which dynamics parameters can be identified and how well simulation reproduces the road. The fixed-price remote diagnosis tells you whether a quick identification routine for each new vehicle is realistic.

Robust planning and control are needed fast for public roads

What happens. A pilot service with driverless vehicles on public roads has to deal with traffic, pedestrians, weather and road works. Planning and control must react well and stay comfortable, and the schedule for the pilot is short.

What it costs. Hard braking and hesitant driving that passengers and traffic notice, safety drivers who intervene often, and a start that slips.

Why the usual fix fails. Rule-based planners grow with every new situation until nobody can follow them. Learned planners are hard to prove safe. Waiting for a full team delays the pilot.

How a fixed-price remote diagnosis checks it. From recorded drives and the current planner and controller I look at where the vehicle brakes or steers harder than needed and which cases cause interventions. The fixed-price remote diagnosis tells you whether predictive planning and control that respects the limits can calm this, and in what order to build it.

Background: Predictive control with obstacle avoidance under uncertainty

A trajectory controller has to be tuned again on each new vehicle platform

What happens. A predictive trajectory controller works on one test vehicle. On the next platform the mass, tyres, actuator delays and steering differ, so it has to be tuned again. Each platform means test days and expert hours.

What it costs. Time for every platform, and a team that is the bottleneck for each new vehicle.

Why the usual fix fails. Copying the old set of parameters gives poor results. Tuning on the track is slow and hard to repeat. Larger tables are the same problem in more cells.

How a fixed-price remote diagnosis checks it. From existing tuning runs on different platforms I look at which parameters follow from vehicle data and which really need testing. The fixed-price remote diagnosis tells you whether model-based tuning can bring a new platform to a good starting point.

Background: Calibration-efficient predictive control