betaFeedback

WorksComputer Science

Old-School Planners Drive (Sometimes) Better Than Learned Ones

Abstract

End-to-end driving models learn to plan from data, and hand-written planners are widely treated as a thing of the past. We test whether this is justified: we implement sixteen textbook planners, each combining a path source (reference line, A*, hybrid A*, RRT*) with a trajectory method (Frenet lattice, nonlinear optimization, ss–tt search, dynamic window), and evaluate them on three benchmarks: open loop on nuScenes, closed loop on nuPlan with ground-truth perception, and photorealistic closed loop on HUGSIM. On HUGSIM, they replace only the planner of the end-to-end model UniAD and read its frozen perception, without an HD map or other privileged information. On nuScenes and nuPlan, the best textbook planners are competitive (L2 error of 0.735,m against 0.947,m for UniAD's planner; a nuPlan score of 79.3 against 76.1 for the rule-based PDM-Closed on Test14-hard with reactive agents). On HUGSIM, nine of the sixteen score above UniAD's learned planner on the same perception,. The best, a hybrid A* path followed by nonlinear trajectory optimization, raises HUGSIM's driving score (HD-Score) from 32.0 to 41.8. The gain is largest with standing and oncoming cars; against cars that steer at the ego, there is none. In 31 of the 182 runs in which UniAD's planner hits a car, the best textbook planner completes the route on the same perception: these crashes are due to the planner, not the perception.

Introduction

End-to-end driving models such as UniAD [1] and its successors [2, 3, 4, 5] map raw sensor data to a trajectory and learn the planner from human driving logs. Research on these models mostly compares them with other learned planners [5, 6, 7]. Textbook planners, i.e., rule-based methods such as hybrid A∗A^* [8], Frenet lattices [9] and optimization [10, 11] are widely treated as a thing of the past and rarely appear as baselines; NAVSIM [12], for instance, reports constant velocity and an ego-status MLP as its simplest ones. Where rule-based planners do beat learned ones, as PDM-Closed [13] did on nuPlan [14], they receive ground-truth objects and an HD map. The planner inside an end-to-end model, in contrast, sees the world through learned, noisy perception. The closest test on such perception, NeuroNCAP [15], found that a simple braking rule on UniAD’s detections already scores above UniAD’s planner, but tested no complete planner.

Two questions are therefore open. How well do textbook planners drive on today’s benchmarks? And do they still drive well when their input is the noisy perception of an end-to-end model instead of ground-truth objects and an HD map? The second also concerns the model itself: in our runs, UniAD crashes in two thirds of the scenarios of the photorealistic simulator HUGSIM [16], and it is unknown whether its perception or its planner is to blame.

We study sixteen textbook planners, each combining a path source with a trajectory method, on three benchmarks of increasing realism: open loop on nuScenes [17] with UniAD’s perception, closed loop on nuPlan with ground-truth perception, and closed loop on HUGSIM with the perception of UniAD. On HUGSIM, we keep the official UniAD checkpoint frozen and replace only its planner (Figure 1). The textbook planners read UniAD’s detected objects, forecasts and road masks, the navigation command and the ego state, use no HD map or other privileged information, and return the same six waypoints to the same controller, so any difference in driving is due to the planner. In summary, we contribute sixteen textbook planners that run on the outputs of an end-to-end model, without an HD map, and show:

Figure 1. (a) On HUGSIM, UniAD’s planner and sixteen textbook planners read the outputs of the same frozen UniAD perception and return six waypoints to the same controller. (b) A car stands still 20 m ahead (numbered dots: ego positions over time). UniAD’s learned planner keeps driving and hits it after 4 s (HD-Score 1 out of 100). The best textbook planner, hybrid A* with trajectory optimization (optim), treats the detected car as an obstacle and passes it on the left (HD-Score 99).

  • On nuScenes and nuPlan, the best textbook planners are competitive: a Frenet lattice reaches an open-loop L2 error of 0.735 m against 0.947 m for UniAD’s planner and comes within 3 points of PDM-Closed on nuPlan (Sec. V) on Val14 and outperforms it on Test14-hard.

  • On UniAD’s perception, nine of the sixteen drive better than UniAD’s learned planner. The best raises the HD-Score from 32.0 to 41.8, above all but one of the learned models evaluated on the same scenarios (Sec. VI).

  • The gain is largest with standing cars, which the path source drives around, and with oncoming cars, which the trajectory method checks against their forecasts. Against cars that steer at the ego, there is none.

End-to-end driving. UniAD [1] chains modules for tracking, online mapping, motion forecasting and occupancy prediction; a learned planner attends to their features, and a post-optimizer pushes the waypoints away from predicted occupancy. Follow-up work changes the scene representation (VAD [2], SparseDrive [4]), runs the modules in parallel (PARA-Drive [3]), generates trajectories by diffusion (DiffusionDrive [5]), or scores many proposals with heads distilled from human and rule-based teachers (Hydra-MDP [6]). NMP, P3 and MP3 [18, 19, 20] select sampled trajectories with learned costs, and SafetyNet [21] and InterFuser [22] check or constrain a learned planner’s output with hand-written rules. In all of these, the component that decides the trajectory is learned or only checks a learned one; we replace it with complete textbook planners.

Why learned planners fail. Behavior cloning needs synthesized perturbations to drive robustly [23] and suffers from dataset bias [24]. Learned planners also pick up shortcuts: they copy the expert’s previous action [25], rely on the ego status [26], [27] or depend on the target point {citep}[28]`.

Rule-based planning and closed-loop benchmarks. On nuPlan, PDM-Closed [13] scores IDM [29] proposals with a function that closely resembles the benchmark metric; the strongest recent planners combine learning with rule-based selection or refinement [30, 31]. All of these use ground-truth perception. NAVSIM [12] evaluates end-to-end models in non-reactive simulation, Bench2Drive [32] fully closed loop in CARLA, and HUGSIM [16] in photorealistic closed loop on reconstructed real drives, where several end-to-end models have been evaluated zero-shot [33, 34].

Sixteen Textbook Planners

Every planner follows the textbook split of motion planning [35]: a path source decides where to drive, and a trajectory method decides how fast to drive along that path and how far to leave it. We implement four of each and combine them, which gives sixteen planners (Figure 2).

Figure 2. Overview of path sources and trajectory methods on a synthetic street (black: ego; gray: other cars; green star: goal). Top, path sources around a parked car: (a) reference line, which follows the road and runs through the parked car (red cross); avoiding it is left to the trajectory method; (b) A* on a grid with blocked cells shaded; (c) hybrid A* with bounded-curvature arcs; (d) RRT*. Bottom, trajectory methods behind a slow car (dots: future ego positions; dashed red: colliding candidates): (e) lattice; (f) optimization from several initial guesses; (g) s–t search; (h) dynamic window.

What the planners read from UniAD

In every cycle, a planner receives UniAD’s tracked agents with motion forecasts over 6 s, the predicted vehicle occupancy, bird’s-eye-view masks of the drivable area, the ego speed, acceleration and yaw rate, and the navigation command (left, right, straight). It returns six waypoints at 0.5 s spacing, the format of UniAD’s own plan.

Map-based planners take the route, goal, free space and speed limit from an HD map; we take them from UniAD’s outputs. A reference line replaces the route: we fan out candidate curves from the ego and take the one that stays longest on the predicted drivable mask among those that turn as the command says. The goal of all path sources is the point on the reference line just beyond what the ego can reach in 3 s, at most 50 m ahead. The predicted drivable mask replaces the free space, and, without a speed limit, the target speed is derived from the current speed, with a tuned minimum.

Static agents become obstacles on the path source’s grid, so the path goes around them. Moving agents are left to the trajectory method, and so is a stopped vehicle that UniAD detects centered in the ego’s lane and aligned with it, which may only be waiting in traffic. All trajectory methods compare the planned ego position at each time step with the forecast position of each agent at that time.

Path sources: where to drive

The four path sources differ in how they get to the goal (Figure 2a–Figure 2d). The simplest uses the reference line itself as the path and ignores obstacles.

A* [36] searches a 0.5 m grid for the cheapest path of the ego’s center. Cells off the drivable mask or within half the ego width of a static obstacle are blocked.

Hybrid A* [8] extends A* to paths the car can drive: it searches over positions and headings with short arcs of bounded curvature and checks the full car footprint against obstacles.

RRT* [37] grows and rewires a tree of random samples, biased towards the reference line, with a minimum turning radius on its edges.

Trajectory methods: when to be where

The four trajectory methods are shown in Figure 2e–Figure 2h. The Frenet lattice planner (lattice) [9] samples polynomials in path coordinates from the measured state to several lateral offsets, durations and end speeds, and adds stopping and following candidates. It returns the cheapest collision-free candidate; the cost covers comfort, speed and distance to the path, plus penalties for proximity to agents and for leaving the drivable area.

The optimization planner (optim) optimizes the controls of a kinematic bicycle model over a 3 s horizon, starting at the measured speed:

min⁡x,u ∑kJks.t.xk+1=f(xk,uk).(1)\min_{x,u}\ \sum_k J_k \quad\text{s.t.}\quad x_{k+1}=f(x_k,u_k). \tag*{(1)}

The running cost JkJ_k has the same terms as the lattice’s. Because the problem is non-convex, we run the gradient-based solver from several initial guesses (the path, copies shifted to the left and right, and a braking profile) and return the collision-free solution with the lowest cost.

The ss–tt search [11] keeps the path fixed and chooses only how far along it the ego travels over time. It works on a distance–time graph, a grid of the distance ss along the path against the time tt. For every agent, it blocks the cells in which the ego would come too close to that agent’s forecast, and then searches for the cheapest speed profile through the remaining cells. Depending on where the blocked cells lie, this profile follows the agent, yields to it, passes before it, or stops.

The dynamic window approach (DWA) [38] considers the speeds and yaw rates that the car can reach within the next 0.5 s and predicts the arc it would drive with each. It discards every arc on which the car could no longer stop before an obstacle and picks the one that best balances heading towards the path, distance to obstacles and speed.

Benchmarks and Tuning

nuScenes. On the validation scenes of nuScenes [17], the planners run open loop on UniAD’s perception. We report two common metrics. The metric of UniAD takes the L2L2 error to the human trajectory at 1, 2 and 3 s and tests collisions with a coarse, axis-aligned box. The metric of Li et al. [27] averages the error up to each horizon, which gives smaller values, and tests the true ego box at every step. Values from the two are not comparable.

nuPlan. On nuPlan [14], the planners run closed loop on ground-truth objects with constant-velocity forecasts and the HD map, whose route centerline replaces the reference line (path source route). We report the closed-loop score on Val14 [13] and Test14-hard [26] with non-reactive (NR) and reactive (R) agents.

HUGSIM. HUGSIM [16] renders camera images from Gaussian-splatting reconstructions of real drives (nuScenes, Waymo, PandaSet, KITTI-360) and inserts cars that stand, drive in the same direction, come towards the ego, cross or steer at the ego. Its 436 scenarios fall into four tiers, from easy (no inserted car) to extreme (mostly cars that steer at the ego). A run starts at 1 m/s and ends at the first collision, when the ego leaves or completes the route, or after about 100 s. The HD-Score (0–100) is the route completion times the mean per-step score for collisions, drivable area, time to collision and comfort. We average it with equal weight for each source dataset and tier, and report gains over UniAD’s planner as per-scenario differences, averaged the same way, ± one standard error. We use the official UniAD v2.0 checkpoint. In our setup, UniAD’s planner, i.e., its final plan after its own collision optimizer, scores 32.0, in line with an independent evaluation [33]. For the textbook planners, UniAD still runs completely on every frame, and only its final plan is replaced.

Parameter tuning. We tune the cost weights and clearance margins of each planner by random search: on nuScenes, on 120 training scenes against the sum of mean L2L2 and mean collision rate; on nuPlan, on 84 training scenarios against the official non-reactive score. HUGSIM received no tuning: the nuScenes configurations were copied unchanged.

plannerUniAD metricmetric of Li et al.
L2ColL2Col
UniAD v1.0 [1], [27]1.030.310.660.62
ReAL-AD [39]0.740.17——
Ego-MLP [27]——0.350.37
UniAD’s planner (v2.0, our run)0.9470.2770.6480.534
reference line + lattice (ours)0.7350.7530.4300.137
hybrid A* + optim (ours)0.8710.7920.4980.241

Table 1. nuScenes, open loop: L2 and collision rate. Bold: best.

plannerVal14Test14-hard
NRRNRR
PDM-Closed [13] (rule-based)92.892.165.175.2
PLUTO [30] (learned + rules)92.989.7^a80.176.9
Diff. Planner + refine [31] (learned + rules)94.392.978.982.0
Diff. Planner [31] (learned)89.982.876.069.2
CaRL [40] (learned, RL)91.390.6——
route + lattice (ours)90.090.475.379.3
hybrid A* + optim (ours)81.684.065.672.9

Table 2. nuPlan closed-loop score with ground-truth perception, with non-reactive (NR) and reactive (R) agents. Bold: best.

Published values from [31]; CaRL and the value marked a^{a} from [40].

Results on nuScenes and nuPlan

nuScenes: low L2, few collisions. On UniAD’s metric, a lattice on the reference line reaches an L2 error of 0.735 m against 0.947 m for UniAD’s planner (Table I). This matches ReAL-AD [40] (0.74 m), a recent learned model that builds on UniAD. Part of this is expected: the textbook planners start from the measured speed and acceleration, which open-loop L2 rewards [27, 41], and learned planners that receive the ego status reach an even lower L2 (0.35 m against 0.43 m on the metric of Li et al.). The collision rate depends on the metric. On the stricter test of Li et al., the same planner collides less often than UniAD’s planner (0.14 % against 0.53 %). On UniAD’s coarse test, it collides more often (0.75 % against 0.28 %); most of these collisions come from two scenes with a standing ego, where the coarse test misplaces the ego box. Across the sixteen planners, the trajectory method matters more than the path source: the lattice has the lowest errors, and DWA, which rolls out constant arcs, the highest.

path sourcetrajectoryUniAD L2Li et al. L2Li et al. ColnuPlan Val14nuPlan T14-hard
UniAD’s planner0.9470.6480.53––
reference linelattice0.7350.4300.1490.079.3
reference lineoptim0.8930.5070.2786.674.8
reference liness–tt0.9490.5400.2682.470.4
reference lineDWA1.2730.7820.5387.776.8
A*lattice0.8320.5070.2476.062.6
A*optim0.9120.5330.2581.068.2
A*ss–tt1.1400.6940.2475.062.8
A*DWA1.3410.8170.5387.276.6
hybrid A*lattice0.7830.4810.1878.566.5
hybrid A*optim0.8710.4980.2481.672.9
hybrid A*ss–tt1.1560.6970.2774.660.7
hybrid A*DWA1.5280.9360.5283.374.7
RRT*lattice0.7900.4670.1582.875.9
RRT*optim0.8580.4900.2381.174.8
RRT*ss–tt1.2260.7050.2383.670.6
RRT*DWA1.3500.8310.4683.474.8

Table 3. All sixteen planners on nuScenes (metrics of UniAD and of Li et al.: L2, collision rate) and nuPlan (Val14 NR, Test14-hard R). Bold: best.

path sourcelatticeoptimss–tt searchDWA
reference line34.9 (+2.8)35.6 (+3.6)29.5 (−2.5)31.5 (−0.5)
A*35.9 (+3.9)33.8 (+1.8)26.2 (−5.8)35.7 (+3.7)
hybrid A*35.1 (+3.1)41.8 (+9.7)30.7 (−1.3)33.9 (+1.9)
RRT*30.7 (−1.3)41.0 (+9.0)31.2 (−0.8)<u>28.0 (−4.0)</u>

Table 4. HUGSIM HD-score of the sixteen textbook planners on UniAD’s perception, with the gain over UniAD’s planner (32.0) in brackets. The standard error of each gain is 1.4–1.6; bold: gain above two standard errors, underlined: loss beyond two.

nuPlan: close to PDM-Closed. With ground-truth objects and the map route, a lattice on the route scores 90.0 on Val14, 2.8 points below PDM-Closed (Table II). On Test14-hard with reactive agents, it scores 79.3, second only to Diffusion Planner with refinement among the planners in Table II. Test14-hard was built from scenarios that are hard for PDM-Closed [26]. Among the path sources, the route is best for every trajectory method except the ss–tt search (Table III): searched paths cut corners through the oncoming lane, which nuPlan penalizes. Unlike PDM-Closed, our planners select trajectories with their own costs, not with a copy of the benchmark metric. No planner is best everywhere: hybrid A* + optim, the best on HUGSIM, reaches 0.871 m on nuScenes and 81.6 on Val14.

methodeasymediumhardextrememean
same 436 scenarios and simulator
hybrid A + optim (ours)*71.048.334.813.041.8
UniAD’s planner (our run)63.233.521.79.632.0
WA-JEPA [33]79.855.630.613.644.9
DrivOR [33, 42]78.029.120.014.135.3
UniAD [1, 33]64.037.221.06.332.1
LTF [12, 33]66.115.512.011.726.3
VAD [2, 33]42.08.57.76.316.1
other simulator versions or scenario sets (indicative)
TerraTransfer++ [43]82.171.249.316.154.7
DriveZero-Scale [34]81.960.524.727.648.7
DriveZero [34]78.547.216.028.742.6
GigaPixel [34]67.451.919.121.640.0
DrivOR-Scale [34]64.850.316.825.839.4

Table 5. HUGSIM HD-score by tier, compared with published methods. Mean: unweighted mean of the four tiers, recomputed by us.

Results on HUGSIM with UniAD’s Perception

Does a textbook planner drive better? Sometimes. On the same perception, nine of the sixteen planners score above UniAD’s planner and seven below (Table IV). Six of the gains and two of the losses exceed two standard errors. The two largest gains are both optim on a searched path: hybrid A* + optim gains 9.7 ± 1.4 points and RRT* + optim 9.0 ± 1.4, so the result is not one lucky configuration. The combination matters: optim is clearly ahead only on the two searched paths with bounded curvature (hybrid A*, RRT*). The ss–tt search is weakest (29.4 on average): with its tuned weights, it barely accelerates from the start speed and completes little of the route. All sixteen planners collide less often per driven step than UniAD’s planner, but only the two best also complete more of the route; the others drive more slowly and get less far.

How does it compare with published models? Table V lists published HUGSIM results. On the same 436 scenarios, as evaluated in [33], the best textbook planner (41.8) is ahead of VAD, LTF, UniAD (32.1) and DrivOR (35.3), and behind only the world-action model WA-JEPA (44.9). Results from other simulator versions differ in scenario set and scoring: UniAD alone scores between 25.7 and 34.5 across them [34, 43]. They are therefore only indicative: the best textbook planner is on par with DriveZero (42.6), ahead of GigaPixel and DrivOR-Scale, and behind DriveZero-Scale, trained with additional simulated data, and TerraTransfer++, which adapts to HUGSIM’s renders.

Where does the gain come from? We now compare the best planner, hybrid A* + optim, with UniAD’s planner (Table VI). The gain is largest with standing and oncoming cars, the situations that the two stages of a textbook planner are built for. Unless UniAD detects it straight ahead in the lane, a standing car is an obstacle on the path source’s grid, so the path goes around it (Figure 1b, Fig. 3b}): UniAD’s planner hits the standing car in 31 of 80 scenarios, the textbook planner in 13. A moving car is checked against its forecast at every time step, so the trajectory makes room, waits or follows (Fig. 3a,c). The tiers show the same pattern (Table V): the gain is largest in the medium and hard tiers (+14.8 and +13.1), which contain the standing, slow and oncoming cars, and small in the extreme tier (+3.4), where most cars steer at the ego. The gain also holds on all four source datasets: +11.9, +12.7 and +10.8 on nuScenes, Waymo and PandaSet scenes, and +6.5 on KITTI-360, where both planners score low.

inserted carUniAD’s plannerhybrid A* + optimgain
standing26.645.4+18.8±4.2+18.8\pm4.2
oncoming12.530.3+17.8±4.1+17.8\pm4.1
crossing22.037.1+15.1±8.0+15.1\pm8.0
same direction44.054.9+10.8±3.4+10.8\pm3.4
several types21.130.9+9.8±5.4+9.8\pm5.4
none59.367.4+8.1±3.0+8.1\pm3.0
steers at ego13.912.9−1.0±2.8-1.0\pm2.8

Table VI. HUGSIM BY TYPE OF INSERTED CAR: PLAIN MEAN OF THE HD-SCORE OVER THE SCENARIOS OF EACH TYPE (HENCE SLIGHTLY DIFFERENT FROM TABLES IV AND V), AND GAIN ± ONE STANDARD ERROR (BOLD: ABOVE TWO STANDARD ERRORS).

Perception or planner? UniAD’s planner hits a car in 182 runs. In 31 of them, the textbook planner completes the route on the same perception: these crashes are due to the planner, not the perception. In another 131, both planners hit a car, and we cannot tell which part is to blame.

Is the gain just caution? No. The textbook planner starts more slowly: from the common start at 1 m/s, it reaches 2.2 m/s after 1 s against 3.5 m/s for UniAD’s planner, which crashes within the first 4 s in 125 runs against 70. But a slow start alone does not explain the gain: it is smallest in the 133 scenarios where the textbook planner is more than 1.5 m/s slower than UniAD’s planner over the first 4 s (+4.0) and largest in the 110 where the two speeds differ by less than 0.5 m/s (+14.6).

Where does textbook planning lose? Against cars that steer at the ego, there is no gain (Table VI). UniAD’s forecasts do not anticipate that such a car turns towards the ego, so the outcome is a matter of timing: in the largest losses, UniAD’s faster planner has already passed when the car arrives, and the slower textbook planner is hit (Figure 3d). The textbook planner also leaves the route more often in turns. Without an HD map, it depends on UniAD’s predicted road masks, which often miss the branch that the route takes.

Figure 3. Four HUGSIM scenarios on the same UniAD perception: bird’s-eye view of both runs (UniAD’s planner orange, dashed; hybrid A* + optim blue) and the front camera at the decisive moment, with the HD-Score of each run. (a) A car approaches in the ego lane; UniAD’s planner, faster, hits it head-on, while the textbook planner arrives later and further right. (b) A car stands just past a right turn; UniAD’s planner turns late and hits it. (c) A car drives ahead at 2 m/s; UniAD’s planner rear-ends it, the textbook planner follows. (d) A car steers at the ego; UniAD’s planner accelerates away, and the slower textbook planner is hit.

Discussion

Implications. Textbook planners are not a thing of the past: the best are competitive on nuScenes and nuPlan and, on the perception of an end-to-end model, drive better than its learned planner. For end-to-end models with interpretable outputs, a textbook planner on the model’s own perception is a cheap and informative baseline: because both planners read the same perception, it separates errors of the planner from errors of the perception. The failure modes are partly complementary, so learned proposals could be checked with an explicit, time-indexed collision test, as rule-based selection does on nuPlan [13, 30].

Limitations. All results on learned perception use one end-to-end model; models with stronger learned planners may behave differently. We did not run the planners on the simulator’s true boxes, so the runs in which both planners hit a car cannot be divided between perception and planning. HUGSIM’s scenarios are variants of a fixed set of base scenes and thus not independent, as our standard errors assume. HUGSIM inserts only cars, and its score mostly judges the planned waypoints, not the driven motion. It also ignores the driving direction, so using the oncoming lane costs nothing.

Conclusion. Old-school planners sometimes drive better than learned ones. The best textbook planners are competitive on nuScenes and nuPlan, and on UniAD’s unchanged perception the best of them clearly beats UniAD’s planner in closed loop. They still fail in interactive scenarios: against cars that steer at the ego, where the forecasts they rely on are wrong, they gain nothing. Combining rule-based and learned planning is therefore a promising direction, as the strongest planners on nuPlan already show [30, 31]. Repeating the substitution with other end-to-end models, and with the simulator’s true boxes to separate perception from planning errors, is future work.

References

References

  1. [1]Y. Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang, L. Lu, X. Jia, Q. Liu, J. Dai, Y. Qiao, and H. Li, “Planning-oriented autonomous driving,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023.arxiv.org/abs/2212.10156
  2. [2]B. Jiang, S. Chen, Q. Xu, B. Liao, J. Chen, H. Zhou, Q. Zhang, W. Liu, C. Huang, and X. Wang, “VAD: Vectorized scene representation for efficient autonomous driving,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023.arxiv.org/abs/2303.12077
  3. [3]X. Weng, B. Ivanovic, Y. Wang, Y. Wang, and M. Pavone, “PARA-Drive: Parallelized architecture for real-time autonomous driving,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024.DOI
  4. [4]W. Sun, X. Lin, Y. Shi, C. Zhang, H. Wu, and S. Zheng, “SparseDrive: End-to-end autonomous driving via sparse scene representation,” 2024, arXiv preprint arXiv:2405.19620.arxiv.org/abs/2405.19620
  5. [5]B. Liao, S. Chen, H. Yin, B. Jiang, C. Wang, S. Yan, X. Zhang, X. Li, Y. Zhang, Q. Zhang, and X. Wang, “DiffusionDrive: Truncated diffusion model for end-to-end autonomous driving,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, arXiv:2411.15139.arxiv.org/abs/2411.15139
  6. [6]Z. Li, K. Li, S. Wang, et al., “Hydra-MDP: End-to-end multimodal planning with multi-target hydra-distillation,” arXiv:2406.06978, 2024.arxiv.org/abs/2406.06978
  7. [7]Y. Li, Y. Wang, Y. Liu, J. He, L. Fan, and Z. Zhang, “End-to-end driving with online trajectory evaluation via BEV world model,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2025, pp. 27137–27146.arxiv.org/abs/2504.01941
  8. [8]D. Dolgov, S. Thrun, M. Montemerlo, and J. Diebel, “Path planning for autonomous vehicles in unknown semi-structured environments,” Int. J. Robot. Res., vol. 29, no. 5, pp. 485–501, 2010.DOI
  9. [9]M. Werling, J. Ziegler, S. Kammel, and S. Thrun, “Optimal trajectory generation for dynamic street scenarios in a frenet frame,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2010.DOI
  10. [10]J. Ziegler, P. Bender, T. Dang, and C. Stiller, “Trajectory planning for Bertha – a local, continuous method,” in Proc. IEEE Intell. Veh. Symp. (IV), 2014.DOI
  11. [11]H. Fan, F. Zhu, C. Liu, L. Zhang, L. Zhuang, D. Li, W. Zhu, J. Hu, H. Li, and Q. Kong, “Baidu apollo EM motion planner,” arXiv preprint arXiv:1807.08048, 2018.arxiv.org/abs/1807.08048
  12. [12]D. Dauner et al., “NAVSIM: Data-driven non-reactive autonomous vehicle simulation and benchmarking,” in Adv. Neural Inf. Process. Syst. (NeurIPS), Datasets and Benchmarks Track, 2024.arxiv.org/abs/2406.15349
  13. [13]D. Dauner, M. Hallgarten, A. Geiger, and K. Chitta, “Parting with misconceptions about learning-based vehicle motion planning,” in Proc. Conf. Robot Learn. (CoRL), 2023.arxiv.org/abs/2306.07962
  14. [14]H. Caesar, J. Kabzan, K. S. Tan, W. K. Fong, E. Wolff, A. Lang, L. Fletcher, O. Beijbom, and S. Omari, “nuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles,” arXiv preprint arXiv:2106.11810, 2021.
  15. [15]W. Ljungbergh et al., “NeuronCAP: Photorealistic closed-loop safety testing for autonomous driving,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2024.arxiv.org/abs/2404.07762
  16. [16]H. Zhou, L. Lin, J. Wang, Y. Lu, D. Bai, B. Liu, Y. Wang, A. Geiger, and Y. Liao, “HUGSIM: A real-time, photo-realistic and closed-loop simulator for autonomous driving,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 48, no. 4, pp. 4673–4691, 2026.DOI
  17. [17]H. H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020.
  18. [18]W. Zeng, W. Luo, S. Suo, A. Sadat, B. Yang, S. Casas, and R. Urtasun, “End-to-end interpretable neural motion planner,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 8660–8669.arxiv.org/abs/2101.06679
  19. [19]A. Sadat, S. Casas, M. Ren, X. Wu, P. Dhawan, and R. Urtasun, “Perceive, predict, and plan: Safe motion planning through interpretable semantic representations,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2020.arxiv.org/abs/2008.05930
  20. [20]S. Casas, A. Sadat, and R. Urtasun, “MP3: A unified model to map, perceive, predict, and plan,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 14403–14412.arxiv.org/abs/2101.06806
  21. [21]M. Vitelli, Y. Chang, Y. Ye, M. Wołczyk, B. Osinskí, M. Niendorf, H. Grimmett, Q. Huang, A. Jain, and P. Ondruska, “SafetyNet: Safe planning for real-world self-driving vehicles using machine-learned policies,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2022.arxiv.org/abs/2109.13602
  22. [22]H. Shao, L. Wang, R. Chen, H. Li, and Y. Liu, “Safety-enhanced autonomous driving using interpretable sensor fusion transformer,” in Proc. Conf. Robot Learn. (CoRL), ser. Proceedings of Machine Learning Research, vol. 205, 2025, pp. 726–737.
  23. [23]M. Bansal, A. Krizhevsky, and A. Ogale, “ChauffeurNet: Learning to drive by imitating the best and synthesizing the worst,” in Proc. Robot.: Sci. Syst. (RSS), 2019.arxiv.org/abs/1812.03079
  24. [24]F. Codevilla, E. Santana, A. M. López, and A. Gaidon, “Exploring the limitations of behavior cloning for autonomous driving,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019.arxiv.org/abs/1904.08980
  25. [25]C. Wen, J. Lin, T. Darrell, D. Jayaraman, and Y. Gao, “Fighting copycat agents in behavioral cloning from observation histories,” in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020.arxiv.org/abs/2010.14876
  26. [26]J. Cheng, Y. Chen, X. Mei, B. Yang, B. Li, and M. Liu, “Rethinking imitation-based planner for autonomous driving,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2024, arXiv:2309.10443.arxiv.org/abs/2309.10443
  27. [27]Z. Li, Z. Yu, S. Lan, J. Li, J. Kautz, T. Lu, and J. M. Alvarez, “Is ego status all you need for open-loop end-to-end autonomous driving?” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024.arxiv.org/abs/2312.03031
  28. [28]B. Jaeger, K. Chitta, and A. Geiger, “Hidden biases of end-to-end driving models,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023.arxiv.org/abs/2306.07957
  29. [29]M. Treiber, A. Hennecke, and D. Helbing, “Congested traffic states in empirical observations and microscopic simulations,” Phys. Rev. E, vol. 62, no. 2, pp. 1805–1824, 2000.DOI
  30. [30]J. Cheng, Y. Chen, and Q. Chen, “PLUTO: Pushing the limit of imitation learning-based planning for autonomous driving,” 2024, arXiv preprint arXiv:2404.14327.arxiv.org/abs/2404.14327
  31. [31]Y. Zheng, R. Liang, K. Zheng, J. Zheng, L. Mao, J. Li, W. Gu, R. Ai, S. E. Li, X. Zhan, and J. Liu, “Diffusion-based planning for autonomous driving with flexible guidance,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2025, arXiv:2501.15564.arxiv.org/abs/2501.15564
  32. [32]X. Jia et al., “Bench2Drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving,” in Adv. Neural Inf. Process. Syst. (NeurIPS), Datasets Benchmarks Track, 2024.arxiv.org/abs/2406.03877
  33. [33]X. Wang, Y. Xiang, Y. Zhou, et al., “WA-JEPA: Rethinking the video JEPA paradigm for world-action modeling in autonomous driving,” 2026, arXiv:2608.20974.arxiv.org/abs/2608.20974
  34. [34]H. He, C. Hu, Z. Su, H. Zhang, H. Liu, J. Li, H. Tian, Z. Shen, H. Li, Z. Li, Y. Yang, B. Huang, S. Zhang, K. Chen, X. Zhang, W. Dai, H. Dai, S. Liu, Z. Huang, and N. Wang, “DriveZero: End-to-end driving beyond human demonstrations,” 2026, arXiv preprint arXiv:2609.06055.arxiv.org/abs/2609.06055
  35. [35]B. Paden, M. Cáp, S. Z. Yong, D. Yershov, and E. Frazzoli, “A survey of motion planning and control techniques for self-driving urban vehicles,” IEEE Trans. Intell. Veh., vol. 1, no. 1, 2016.arxiv.org/abs/1604.07446
  36. [36]P. E. Hart, N. J. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,” IEEE Trans. Syst. Sci. Cybern., vol. 4, no. 2, pp. 100–107, 1968.DOI
  37. [37]S. Karaman and E. Frazzoli, “Sampling-based algorithms for optimal motion planning,” Int. J. Robot. Res., vol. 30, no. 7, pp. 846–894, 2011.arxiv.org/abs/1105.1186
  38. [38]D. Fox, W. Burgard, and S. Thrun, “The dynamic window approach to collision avoidance,” IEEE Robot. Autom. Mag., vol. 4, no. 1, pp. 23–33, 1997.DOI
  39. [39]Y. Lu, J. Tu, Y. Ma, and X. Zhu, “ReAL-AD: Towards human-like reasoning in end-to-end autonomous driving,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2025.arxiv.org/abs/2507.12499
  40. [40]B. Jaeger, D. Dauner, J. Beißwenger, S. Gerstenecker, K. Chitta, and A. Geiger, “CaRL: Learning scalable planning policies with simple rewards,” in Proc. Conf. Robot Learn. (CoRL), ser. PMLR, vol. 305, 2025, pp. 5301–5338.arxiv.org/abs/2504.17838
  41. [41]J.-T. Zhai, Z. Feng, J. Du, Y. Mao, J.-J. Liu, Z. Tan, Y. Zhang, X. Ye, and J. Wang, “Rethinking the open-loop evaluation of end-to-end autonomous driving in nuScenes,” 2023, arXiv preprint arXiv:2305.10430.arxiv.org/abs/2305.10430
  42. [42]E. Kirby, A. Boulch, Y. Xu, Y. Yin, G. Puy, É. Zablocki, A. Bursuc, S. Gidaris, R. Marlet, F. Bartoccioni, A.-Q. Cao, N. Samet, T.-H. Vu, and M. Cord, “Driving on registers,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026, pp. 32058–32069.arxiv.org/abs/2601.05083
  43. [43]Z. Xiong, W. Li, Z. Wu, A. Rangesh, S. Bonde, G. Hall, C. Tang, Y. Hu, and W. Zhan, “TerraTransformer: Demonstration-free end-to-end planner,” 2026, arXiv preprint arXiv:2610.30319.arxiv.org/abs/2610.30319

Paper details

vlnc:W_d6spckpb
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents