Hidden Biases of End-to-End Driving Models

Bernhard Jaeger; Kashyap Chitta; Andreas Geiger

Source-linked AI summary

Hidden Biases of End-to-End Driving Models

Bernhard Jaeger, Kashyap Chitta, Andreas Geiger

arXiv:2306.07957v2Jun 13, 2023cs.CVcs.AIcs.LGcs.RO

TL;DR

End-to-end driving progress on CARLA is difficult to attribute because methods and training sets vary widely. This paper identifies recurring target-point and waypoint biases, develops TF++, and achieves state-of-the-art performance on Longest6 and LAV.

Problem
The sources of rapid CARLA performance gains remain unclear, while target-point recovery and waypoint outputs introduce shortcuts and ambiguity.
Method
The paper uses controlled experiments to analyze target-point and waypoint biases, then develops TF++ with alternative path-based outputs and uncertainty-aware speed prediction.
Results
TF++ places first on the Longest6 and LAV benchmarks and outperforms Interfuser by a large margin on Longest6.
Takeaways & Limitations
Target-point recovery and waypoint interpolation are important recurring behaviors in current CARLA driving systems, motivating simpler alternatives that improve benchmark performance.

Abstract

from arXiv · show

End-to-end driving systems have recently made rapid progress, in particular on CARLA. Independent of their major contribution, they introduce changes to minor system components. Consequently, the source of improvements is unclear. We identify two biases that recur in nearly all state-of-the-art methods and are critical for the observed progress on CARLA: (1) lateral recovery via a strong inductive bias towards target point following, and (2) longitudinal averaging of multimodal waypoint predictions for slowing down. We investigate the drawbacks of these biases and identify principled alternatives. By incorporating our insights, we develop TF++, a simple end-to-end method that ranks first on the Longest6 and LAV benchmarks, gaining 11 driving score over the best prior work on Longest6.

1. Introduction

The paper identifies target-point following and waypoint averaging as hidden biases behind recent CARLA progress, then uses controlled experiments to develop TF++, which achieves state-of-the-art benchmark performance.

Motivation: End-to-end driving performance on CARLA rose from under 20 to over 70 in two years, but the reasons remain unclear because methods and training sets differ substantially.Rigorous ablations are costly because driving systems have a large design space and require extensive simulation.
Background: Recent imitation-learning methods achieve high route completion without HD-map inputs by receiving map-based GNSS target points spaced about 30 m apart on average.These target points are located in the lane center and provide geometric information used by current methods.
Hidden biases: Target-point-conditioned models recover from steering errors by steering toward the nearest target point, exploiting map geometry as a shortcut that periodically resets errors.This recovery behavior appears when deployment states fall outside the training distribution and implicitly relies on accurate map information.
Hidden biases: Waypoint outputs are ambiguous because future velocity is multimodal, yet their continuous representation can interpolate between modes and sometimes reduce collisions by slowing down.The paper proposes target-speed classification to represent uncertainty explicitly and interpolate target speeds according to predicted uncertainty.
Contributions: TF++ places first on the Longest6 and LAV benchmarks, outperforming Interfuser by a large margin on Longest6 despite being significantly simpler.The method incorporates insights from controlled experiments and uses approximately four times less computation than the truncated passage specifies.

2. Related Work

Related work traces autonomous-driving imitation learning from early discrete navigation commands toward modern target-point-conditioned waypoint prediction. The paper situates its contributions among LAV, TCP, TransFuser, Interfuser, and concurrent waypoint-based methods.

Imitation learning for autonomous driving: Modern CARLA methods outperform early discrete-navigation-command approaches by predicting waypoints conditioned on target points.Imitation learning dates back over 30 years and regained traction with seminal work and CARLA’s release.
Related waypoint methods: LAV refines navigation-command-conditioned waypoints with a GRU, while the paper attributes much of its steering improvement to target-point conditioning.LAV reports a +50 route-completion improvement from refinement, whose GRU alone receives target-point information.
Related waypoint methods: TCP uses waypoints for stronger collision avoidance than direct control prediction but identifies weaknesses on large turns.This motivates examining waypoint-based representations alongside their limitations.
Recent end-to-end driving approaches: TransFuser is a widely used CARLA baseline, while Interfuser and concurrent methods likewise use structured representations, transformer decoding, target-point conditioning, and waypoint outputs.The paper explains TransFuser’s route-completion gap and proposes architectural, output, and training modifications; Interfuser uses paths, density maps, traffic flags, and heuristic control conversion.

3. Hidden Biases of End-to-End Driving

The section identifies target-point conditioning as a recovery shortcut that explains high route completion but can cause catastrophic steering errors, and shows that pooling spatial features mitigates this failure. It also finds that single point-estimate waypoint predictions inadequately represent multimodal future velocities, causing uncertainty-driven slowing behavior.

Target-Point Shortcut: Target-point conditioning improves recovery, producing a 28-point route-completion advantage over discrete conditioning, but can extrapolate across turns and cause catastrophic steering errors.The target-point model records 0 route deviations per kilometer, whereas discrete conditioning remains unsatisfactory despite augmentation.
Pooling: Replacing global average pooling with spatially preserving transformer decoding increases route completion by 9 points and halves static-object collisions per kilometer.Preserving spatial information helps avoid failures attributed to the target-point shortcut.
Longitudinal Bias: Current methods jointly encode deterministic future paths and multimodal future velocities as waypoints but output only one point estimate, modeling a single mode.The expert uses four target speeds—29, 18, 7, or 0 km/h—while TransFuser predictions span the full 0–29 km/h range.
Longitudinal Bias: When uncertainty lowers the predicted target speed, the car slows down enough to merge successfully behind a cyclist.This qualitative behavior illustrates the longitudinal effect of multimodal speed uncertainty.

4. Comparison to State of the Art

This section compares TF++ with several state-of-the-art driving models and evaluates performance on Longest6 and LAV, alongside runtime measurements. It also describes the multimodal architectures and planning strategies of the comparison methods.

Methods: TF++ is compared against several state-of-the-art models, including Interfuser, TransFuser, and LAV.Interfuser processes multiple cameras and BEV LiDAR with CNNs and transformers; TransFuser fuses camera and LiDAR features; LAV plans from an intermediate BEV representation.
Benchmark evaluation: The evaluation reports performance on training towns in Longest6 and validation towns in LAV.Released models are evaluated with standard deviations over three evaluations on Longest6, while reproduced models report standard deviations over three trainings and three evaluations each on LAV.
Runtime: TF++ outperforms TransFuser by a wide margin in runtime despite using a similar inference compute budget.Runtime is measured per frame in milliseconds, averaged over 300 time steps on one RTX 3090 route.

5. Conclusion · Appendix · A. Changes to TransFuser

The paper attributes strong route-following performance to a learned nearby-target-point bias and presents TransFuser++ as a new state-of-the-art method on LAV and Longest6. It also notes CARLA’s low-speed urban-driving scope, broader societal implications, and implementation differences in the reproduced TransFuser.

5. Conclusion: Recent state-of-the-art driving models achieve exceptional route following by learning a strong bias toward nearby target points.The paper frames this as shortcut learning, extending prior observations of similar biases in autonomous driving.
5. Conclusion: TransFuser++ sets a new state of the art on the LAV and Longest6 benchmarks.The method improves the popular TransFuser baseline through controlled experiments.
5. Conclusion: The study is limited to urban driving in CARLA, where the investigated methods operate below 35km/h.The conclusion identifies high-speed-driving problems as outside the study’s scope.
5. Conclusion: The technology could reduce road accidents and transportation costs while improving mobility for elderly people.These are identified as potential positive societal impacts if the technology is realized.
5. Conclusion: Potential negative implications include reduced employment for human drivers and possible military applications.The paper presents these as societal risks of the technology.
Appendix: The work acknowledges support from research grants, universities, and German government funding programs.The acknowledgements name ERC, BMWi, DFG, BMBF, and IMPRS-IS support.
A. Changes to TransFuser: The reproduced TransFuser largely follows the original implementation but differs in several minor implementation details.These differences are described in the appendix section on changes to TransFuser.

A.1. Expert

The expert driver generates imitation labels using an automatic labeling algorithm and incorporates speed-dependent, road-following safety measures. On Longest6, this expert outperforms the baseline by 4 DS, raising the upper bound for imitation learning models without directly improving current model performance.

A.1. Expert: The expert generates imitation labels with an automatic labeling algorithm, while CARLA supplies auxiliary perception labels except for rendered BEV segmentation.The ego action is approximated with a PID controller following the A⋆ path.
A.1. Expert: Its safety area follows the road and expands with speed; pedestrians within 30 meters trigger preemptive slowing to 2 m/s.The method uses a kinematic bicycle model and sets target speed to 0 when a predicted expert bounding box intersects current agents.
A.1. Expert: 4 DS: The expert outperforms the baseline on Longest6, establishing a higher upper bound for imitation learning models but no direct improvement in current methods.The paper attributes the lack of direct improvement to no method currently reaching the expert’s upper bound.

A.2. Dataset

The training dataset is generated by running the Section A.1 expert on training routes, randomizing weather per route and training at an effective 4 FPS. For the final model, data are recollected at 4 FPS and stored in compressed formats to reduce release size and storage requirements.

Dataset generation: Training data come from every frame of expert-driven training routes, with weather randomized per route and training performed on every fifth frame at an effective 4 FPS.For the scaling experiment, each route is recollected three times, randomizing weather and traffic.
Final dataset: The final model uses data recollected at 4 FPS and trains on every frame, preserving equivalent data while making the released dataset 5x smaller.Camera images, perception labels, text files, and LiDAR point clouds are compressed using JPG, PNG, zip, and LASzip, respectively.

A.3. Training and Architecture … B.1. Attention Pooling Implementation

The implementation combines speed-aware training, filtered localization, disentangled waypoint control, and attention pooling, with a stop-sign controller reducing infractions fivefold and improving driving score by 10.

A.3. Training and Architecture: The target-speed branch uses class-frequency-weighted cross-entropy over four expert speeds with label smoothing 0.1.Label smoothing addresses overconfident predictions that can delay braking behind another vehicle.
A.3. Training and Architecture: Training applies equally weighted losses, AdamW with amsgrad, a 0.0003 learning rate, decay after epoch 30, and 31 epochs.The dense dataset is subsampled fivefold with GPU-indexed frame shifts, and the final model uses four A100 GPUs with batch size 128.
A.3. Training and Architecture: The reproduced TransFuser predicts eight waypoints spaced 250 ms apart, using a single high-resolution camera and LiDAR sensor setup.The camera has a 110° horizontal field of view and 256 x 1024 resolution.
A.4. Localization: Localization uses noisy GNSS measurements filtered by an Unscented Kalman Filter with the expert’s kinematic bicycle model.The GNSS signal has average localization errors of approximately 0.7 meters before filtering.
A.5. PID Controller: The controller uses future waypoint velocity as target speed and reuses the expert’s PID controllers with disentangled waypoint representations.Target speed is computed from the velocity between waypoints 0.5 and 1 second into the future.
A.5. PID Controller: 5x fewer Stop infractions yielded a +10 DS improvement after adding the stop-sign controller.TF++ detects most stop signs, while the controller addresses cases where the stop marking becomes occluded.
B.1. Attention Pooling Implementation: Attention pooling projects 8x8 BEV features from 1512 to 256 channels, adds sinusoidal positional embeddings, and embeds normalized velocity with a two-layer MLP.The velocity embedding is also given a learnable positional encoding.

B.2. Data Augmentation · C. Additional Results

The data augmentation procedure creates a perturbed camera view for every collected frame by shifting and rotating the camera, then trains on it probabilistically with transformed labels. These labels remain centered-lane waypoints, enabling the PID controller to produce lateral recovery toward the lane center.

B.2. Data Augmentation: Every dataset frame receives an augmented counterpart collected with an additional vehicle-mounted camera.
B.2. Data Augmentation: The augmentation shifts the camera ±1 meter laterally and rotates it ±5° around the yaw axis.
B.2. Data Augmentation: The shift and rotation values are sampled from a uniform distribution.
B.2. Data Augmentation: During training, the perturbed camera is loaded with 50% probability.
B.2. Data Augmentation: Waypoint labels and other data are transformed so the perturbed camera becomes the reference center.
B.2. Data Augmentation: The augmented waypoints are original center-lane waypoints rather than expert recovery trajectories, causing the PID controller to steer predictions back toward lane center.

C.1. Longest6 Ablations … C.4. Additional Examples

The supplementary experiments validate TF++ design choices across Longest6, brake-threshold settings, additional baselines, and qualitative recovery examples. They show threshold-dependent safety–completion tradeoffs, generalization effects, representation limitations, and both beneficial and harmful target-point conditioning.

C.1. Longest6 Ablations: The reproduced Longest6 experiments add one TransFuser change per row, averaging three training seeds with three evaluations and reporting training-seed standard deviations.The reproduced baseline’s main problems include high collisions with other vehicles, according to the supplied passage.
C.2. Brake Threshold: Lower brake thresholds reduce vehicle collisions but lower route completion through false-positive braking, with only a 3 DS gap between the best and worst validation settings.The default threshold is 50% for validation towns and 33% for Longest6, where dense traffic emphasizes collision avoidance.
C.3. Additional Baselines: Training on validation towns increases TF++ driving score by 20 points, improving both vehicle collisions and stop sign infractions, though this is not a fair comparison.This experiment illustrates the contribution of generalization issues and brings TF++ close to expert-level performance.
C.3. Additional Baselines: Perception PlanT achieves a DS of 37, but its handcrafted bounding-box representation omits stop signs and incurs a 1.87 stop-sign infraction penalty.The experiment highlights interpretability alongside the risk that human-designed representations miss important scene elements.
C.3. Additional Baselines: Reproduced LAV v2 achieves a driving score of 27, with high stop-sign and red-light infractions concentrated especially in Town 02 under the expanded scenarios.The added scenarios cause many traffic lights to turn yellow and then red as the agent approaches intersections.
C.4. Additional Examples: Target-point conditioning enables TransFuser and TCP to recover from being forced onto a sidewalk, whereas discrete-conditioned TransFuser gets stuck without geometric target-point information.The target-point-conditioned methods extrapolate waypoints toward a nearby target point and return to the lane center.
C.4. Additional Examples: Transformer decoding mitigates harmful target-point extrapolation that makes TCP and TransFuser predict off-road waypoints during a right turn with the target point far behind.The mitigation replaces TransFuser’s global-average-pooling-plus-MLP approach with a transformer decoder.

C.5. Additional Experiments

Additional experiments show that multimodal waypoint prediction is more complex without improving driving score, while target-point conditioning strongly influences behavior across architectures and closely matches steering direction.

Multimodal waypoints: Multimodal waypoints perform 2 DS worse than standard unimodal waypoints, despite using two waypoint GRUs and a selection head.The multimodal representation is more complex, while the reported difference is described as small.
NC conditioned AIM: Target-point conditioning strongly affects route compliance and deviations in AIM, indicating that the target-point bias is architecture-independent.AIM uses a target-point-conditioned waypoint GRU but no LiDAR, auxiliary losses, or transformers; it also performs 13 DS worse than TransFuser overall.
Target Point statistics: 92% of the final model’s validation-route steering signs match the target point’s sign, compared with 93% for the expert.The comparison considers steering angles larger than 1°.

C.6. CARLA Leaderboard

The CARLA leaderboard standardizes evaluation across 20 routes repeated with five random seeds, but released-code results can diverge sharply from reported scores. Its secret evaluation, unreliable auxiliary metrics, absent official validation benchmark, and long submission times limit interpretability and reproducibility.

CARLA leaderboard: The leaderboard evaluates agents on 20 route-weather combinations across 10 routes and 2 secret towns, repeating them five times with different random seeds.Average metrics across all routes are reported to users.
CARLA leaderboard: Released-code evaluation reveals major disparities from reported CARLA leaderboard results, with Interfuser achieving less than half its reported score.These discrepancies suggest that factors not explicitly emphasized in the papers substantially influence leaderboard outcomes.
CARLA leaderboard: TF++ and TF++ WP exhibit a 9 DS difference between representations on the CARLA leaderboard.The submissions used released code and models available around April 2023, and results may change if repositories are updated.
CARLA leaderboard: The leaderboard lacks an official validation benchmark, and existing public-town benchmarks cannot reliably predict leaderboard performance.Available validation benchmarks differ in towns, route length, scenarios, and weather.
CARLA leaderboard: Leaderboard submissions can take up to four weeks because of queues and evaluation time, whereas parallel cluster evaluation of a similar simulation load takes 3–6 hours.Observed pending times reached three weeks, followed by more than four days of evaluation.