Passing a Traction Budget Across a Path–Speed Hierarchy

Many learned driving systems split the job in two. One module draws the path the car should follow, and a second module decides how fast to travel along it. The split is convenient, but it ignores a physical fact: a tire has one fixed amount of grip. Grip spent on turning is not available for braking. The path module does not know how much grip is left, so on wet or icy roads, or in an emergency, it can draw a path the car cannot physically follow at any safe speed. The speed module can only slow down along a path that was already a bad choice.

We propose to make the interface between the two modules carry a traction budget. Given a candidate path and an estimate of road friction, simple physics gives an exact interval of allowed acceleration and braking at every point along that path. It also gives the fastest speed profile the car can actually drive there, and whether the car can still stop before a conflict. The high-level module uses this information to choose among candidate paths by a fixed priority order instead of a weighted reward. The low-level speed controller treats the interval as a hard limit on its actions. This arrangement also fixes a well-documented reward bug. Once progress is measured along a finite path, a single checkable inequality guarantees that crashing always scores worse than standing still.

Core Claim

Conditioning speed on a chosen path, rather than predicting the two in parallel, turns the coupled friction limit into a one-dimensional constraint with a closed form. The same conditioning bounds the total progress reward the car can earn. Together these let a planner decoupled from its controller stay aware of physical limits without either module running at the other's rate. They also remove the manual weighting of safety against progress that makes driving rewards fragile.

Concretely, treat the car as a point mass whose total tire force is limited by the friction circle $\sqrt{F_x^2+F_y^2}\le \mu F_z$. Travelling at speed $v$ along a path with local curvature $\kappa$ uses lateral acceleration $v^2\kappa$. The admissible longitudinal acceleration is then

$$a_x \in \left[-\sqrt{(\mu g)^2-(v^2\kappa)^2},\; \min\!\left(a_{\max},\sqrt{(\mu g)^2-(v^2\kappa)^2}\right)\right], \qquad v \le \sqrt{\mu g/|\kappa|},$$

where $\mu$ is the road friction coefficient, $g$ is gravitational acceleration, and $a_{\max}$ is the powertrain limit. A forward–backward pass over the path turns this pointwise interval into the fastest feasible speed profile. It also yields the shortest stopping distance, with both braking and turning accounted for, to any predicted conflict point.

The speed controller receives a reward that uses only progress along the path:

$$r_t = \frac{\Delta s_t}{L}, \qquad r_{\text{collision}} = -(1+\epsilon).$$

Here $\Delta s_t$ is the arc length covered at step $t$, $L$ is the path length, and $\epsilon>0$. Progress cannot exceed $L$, so every crashing trajectory returns less than $-\epsilon$. Standing still returns $0$ and reaching the goal returns $1$. The ordering that most published driving rewards violate therefore holds by construction. Comfort and physical limits never enter the reward, because they live in the action interval.

Relation to Prior Work

Three papers motivate the gap. AlignDrive predicts a lateral drive path and then predicts one-dimensional displacements along it. This cascade cut collisions from 22.7% to 11.4% in ablation and reached a Driving Score of 89.07 on Bench2Drive. Its output is purely kinematic, however, and nothing stops it from proposing a curve that exceeds available grip. We keep its cascade and add the traction interval and path selection it lacks. SuSIE names the underlying problem as "decoupled awareness": high-level planners imagine subgoals the controller cannot reach. HiRT shows that a slow reasoning module can guide a 10 Hz controller, but only for objects moving at about 1 cm/s. Our setting tests that style of asynchronous guidance where latency meets inertia.

Two papers supply the method, and both handle constraints outside the reward. Action Mapping RL projects racing actions onto the friction boundary. It reports 90% success on Track-A and zero constraint violations, but its boundary is precomputed by discretizing a known model over the full two-dimensional action space, and it has no traffic or planner. BarrierNet filters controls through differentiable barrier functions. It reduces the crash rate from 0.53 to 0.28 in VISTA, but it enforces geometric boundaries, not grip. Our contribution is to observe that once the path is fixed, the action-mapping boundary becomes a closed-form interval that needs no precomputation. That boundary can also be passed upward as a planning signal rather than used only as a filter.

Two papers constrain the reward design. Reward (Mis)design for Autonomous Driving shows that 7 of 9 focus papers prefer crashing to idling, and that lane-centering terms discourage life-saving swerves. A review of driving rewards shows that linear sums let agents trade safety for speed and that fixed weights fail across contexts. Our bounded-progress reward and priority-ordered path choice are direct responses to both findings.

Imminent Collision Mitigation shows that combined braking and steering cuts injury rates by up to 60% relative to straight-line braking when the time to collision is 0.5–1.5 s. We adopt its impact-severity idea as the second-priority criterion. We differ from it by placing that criterion inside a long-horizon hierarchy rather than a monolithic short-horizon policy.

Research Program

The high-level module is AlignDrive's path predictor, modified to emit $K$ candidate paths (for example, keep lane, shift left, shift right, shoulder) at 2–5 Hz. A feasibility evaluator runs at the controller rate of 20–50 Hz and needs no learning. For each candidate, it computes the traction-limited speed profile and checks whether a stop before every predicted conflict is feasible.

Selection follows a strict priority order with no weights. The evaluator first keeps paths that admit a feasible conflict-free profile. If none exist, it chooses the path that minimizes predicted impact speed, weighted by impact angle following Porav and Newman's severity model. Among the paths that remain, it chooses the one with the shortest traction-feasible time to goal. The evaluator also signals the planner to replan early when the current path's feasibility fails between planner ticks. This feedback is the channel HiRT-style systems lack.

The low-level module is a longitudinal policy trained with TD3 or PPO on the bounded-progress reward. It outputs a scalar that is squashed into the current admissible interval, with jerk limits applied as an additional interval. Friction $\mu$ is known in the main experiments and deliberately mis-estimated in a sensitivity sweep. We treat learned tire models of the kind in Autonomous Drifting with 3 Minutes of Data as a later extension, not a dependency.

The analysis has three parts. First, we measure how often unmodified AlignDrive paths are infeasible as friction decreases. Second, we measure how much of the closed-loop gain comes from choosing a different path versus simply slowing down. Third, we verify the crash < idle < goal ordering empirically on logged trajectories using Knox et al.'s sanity checks.

Baselines

The naive reference point is classical longitudinal-only control. It consists of a car-following speed controller plus an independent lateral path tracker on AlignDrive's single path, and it represents industrial practice. A second naive reference is straight-line emergency braking in the collision-mitigation scenarios.

The strongest existing method is AlignDrive itself. It reports a Driving Score of 89.07 and a success rate of 73.18% on Bench2Drive, and its path-conditioned ablation reaches an 11.4% collision rate. We run its released configuration on the official Bench2Drive training set with an identical perception backbone and the same routes. On nominal-friction Bench2Drive our system should not fall below it: we must match its Driving Score within seed variance. On the reduced-friction variants and in emergency scenarios, beating it means a lower collision rate at equal or better route completion. We found no reported result on Lacuna for AlignDrive, or any method, on reduced-friction Bench2Drive, so we establish those numbers ourselves.

Raw2Drive, the strongest reinforcement-learning competitor, reports a 50.24% success rate on Bench2Drive and uses a hand-specified reward head. We cite that number rather than retrain it, because retraining costs 64 H800 GPU days. For a like-for-like reward comparison, we instead train a monolithic PPO driver with a weighted progress, safety, comfort, and lane-centering reward. Its weights are tuned by Bayesian optimization under the same environment-step budget our system receives for all its tuning.

Three ablations isolate the ingredients. The first is AlignDrive plus a post-hoc curvature speed clamp $v\le\sqrt{\mu g/|\kappa|}$. It has friction awareness but no path choice, and it isolates whether selection matters. The second is our system with the priority order replaced by a weighted score over the same three criteria. It isolates whether dropping weights matters. The third is our system with the RL speed controller replaced by a model-predictive speed tracker under the same interval. It isolates whether learning the longitudinal controller adds anything.

For the low level alone, we compare against TD3 with action mapping on Track-A/B. That method reports 90% success on Track-A and a lap time of 1:04.16 on Track-B. Matching it with a boundary computed in closed form, with no precomputation, counts as success.

Why not just…?

Why not just clamp AlignDrive's speed to the curvature limit? This is the most serious threat. On dry roads with gentle curves, friction almost never binds, and the clamp will probably match our system. The clamp cannot help when the chosen path is itself the problem, for example when braking along the lane is insufficient but a lane shift would be feasible. The first ablation measures exactly this. If the clamp matches us on reduced-friction and emergency scenarios as well, the path-selection claim fails, and we would report the contribution as the reward construction alone.

Why not just tune a weighted reward more carefully? Knox et al. found crash-preferring rewards in 7 of 9 carefully engineered papers, and the review shows weights tuned for one context fail in another. Our weighted-PPO baseline gets the same tuning budget. We will check its learned behavior against the crash-versus-idle ordering and evaluate its transfer from nominal to low-friction routes without retuning. The claim is falsified if it passes the ordering check and matches our collision rate.

Why not just use a coupled-tire model-predictive controller for everything? The drifting work shows that such controllers track aggressive maneuvers well, and cooperative platoon controllers of this type coordinate lateral and longitudinal motion. Both still rely on hand-weighted multi-objective costs and have no semantic path proposals. Our third ablation keeps the interface and swaps in model-predictive control at the bottom. If it matches, we conclude that the traction interface is what matters, which is still the core claim, and that reinforcement learning at the low level is optional.

Why not just run a monolithic end-to-end driver? Raw2Drive's roughly 600 ms encoder latency means the car travels about 18 m blind at highway speed. Bench2Drive shows that low open-loop error can coexist with 0% closed-loop success. Our monolithic PPO baseline tests whether implicit coupling learned end to end handles low friction. We expect it to fail through the same reward trade-offs Knox et al. document.

Proposed Experiments

The primary testbed is Bench2Drive's 220 routes, run once at nominal friction and again with tire friction scaled to wet and icy levels. The scaling is applied through the simulator's vehicle physics settings, and the 44 scenario types are kept unchanged. We report Driving Score, success rate, collision rate, efficiency, and comfort.

The second testbed reproduces Porav and Newman's CARLA emergency sweep over initial speed and a time to collision of 0.5–1.5 s. We report injury-weighted impact speed against straight-line braking and against AlignDrive with the clamp. The third testbed is Track-A/B, which checks the low-level controller at the limits of handling.

Three results would be decisive. First, on the low-friction routes and emergency sweep, the gap between our system and the clamp ablation must be significant, which is evidence that the traction budget has to reach the planner and not only the controller. Second, the priority order must beat the weighted variant on transfer from nominal to icy routes, where fixed weights are known to fail. Third, the friction sensitivity sweep must show graceful degradation when $\mu$ is overestimated by 10–30%, which is the realistic failure mode. If the first check fails, the proposal reduces to a reward-design result, and we would say so.