Mapping Driving Policies to Physical Reality

Learned driving policies currently rely on fixed-grid action spaces that permit physically impossible maneuvers. This leads to models that optimize for simulator rewards rather than kinematic viability. The Comfort-by-Construction framework addresses this by dynamically rediscretizing the action grid at each timestep to match the vehicle's instantaneous physical constraints. This approach replaces the common practice of clamping static values, which historically caused grid collapse where control sensitivity vanished as vehicle speeds increased.

The Mechanism of Adaptive Parameterization

Driving simulators typically command accelerations and steering rates as if the vehicle were a point mass unaffected by dynamics. Because these models prioritize cumulative rewards, they learn to exploit gaps in the simulator's physics engine. A policy might perform an abrupt, last-second lane change to avoid an obstacle because the simulator does not explicitly penalize the resulting lateral jerk. When developers force comfort by clipping these values, they shrink the effective action space and leave the model with no meaningful control over its trajectory.

Grid collapse occurs because of how fixed-grid discretization interacts with speed-dependent kinematic limits. In a static grid, the model selects from a predetermined set of values. As vehicle speed increases, the physical limit for a stable, comfortable turn decreases significantly due to tire friction and lateral acceleration bounds. If the static grid range remains fixed while the available physical envelope shrinks, the majority of the model's choices map to physically impossible commands. These are then clipped to the narrow boundary, effectively collapsing multiple distinct policy outputs onto a single point of maximum allowable acceleration. This eliminates the model's ability to distinguish between different control intensities, leading to erratic, low-sensitivity maneuvers.

Adaptive parameterization avoids this by mapping the normalized model output to a physically feasible range. Rather than outputting absolute values, the model outputs a normalized value within a standard range, such as [-1, 1]. The system then multiplies this normalized intent by the dynamic limits computed at that specific timestep. The logic relies on a closed-form inversion of the bicycle model, which relates lateral acceleration to the vehicle’s slip angle and forward velocity. By solving for the steering angle that satisfies both the target acceleration and the friction constraints defined by the tire slip limits, the system defines the boundaries of the action space at each control interval.

Implementation and Boundary Conditions

To see this in practice, consider a hypothetical case where a vehicle is entering a highway off-ramp. Suppose the system determines that the maximum safe lateral acceleration adjusts based on the vehicle’s current trajectory and comfort bounds. If the model outputs a normalized request of 0.9, the system scales this to 90% of the currently allowed acceleration ceiling. If the vehicle were moving faster, the absolute physical ceiling would decrease, but the model's request of 0.9 would still map to 90% of that new, smaller limit. This mechanism ensures the policy maintains a consistent ratio of intent relative to its operational environment, preventing the loss of sensitivity that happens when multiple commands are clipped to the same hard ceiling.

The current framework focuses on internal passenger comfort, but maneuvering around other vehicles presents a conflict. High-jerk maneuvers are sometimes necessary to avoid collisions, yet the ego-vehicle's internal constraint layer may classify these as 'uncomfortable.' Future iterations require an integration of external perception data—specifically predicted occupancy grids from surrounding agents—directly into the inversion model. This would allow the system to dynamically raise or lower the jerk threshold depending on whether the maneuver is a comfort-priority lane change or an safety-priority evasive action. Relying on static thresholds fails here because it cannot distinguish between comfort-induced jerk and collision-avoidance requirements.