Investigating Behavior Cloning from Few Demonstrations for Autonomous Driving based on Bird's-Eye View in Simulated Cities
This paper tackles behavior cloning (BC) for autonomous driving in the CARLA simulator when only a handful of expert demonstrations are available. The core idea is to reweight the supervised BC loss by the inverse of a kernel density estimate over actions, up-weighting rare but safety-critical behaviors such as stopping at red lights and accelerating from a standstill. The motivation is that uniform BC easily ignores these low-density events in small datasets, which can lead to dangerous traffic-light violations and poor generalization.
The paper proposes a sensible and easy-to-implement modification to behavior cloning: using inverse-KDE weights to compensate for severe action imbalance in small expert datasets. The gains over vanilla BC are large and consistent, but the overall contribution is narrow. The experimental evidence rests on a tiny dataset of eight routes, the positioning against prior work is misleading because Chauffeurnet and Codevilla et al. already study traffic-light-aware driving, and critical implementation details such as KDE bandwidth and training hyperparameters are omitted. The promising few-shot results merit attention, yet the paper does not establish whether the method scales beyond a handful of fixed routes or outperforms alternative rebalancing strategies.
The motivation for reweighting is well grounded. Figure 6 and the accompanying analysis show that traffic-light transitions create genuinely rare action clusters that are easily swamped by straight-road data, lending credibility to the claim that uniform BC fails on small datasets. The use of a structured bird's-eye view is pragmatic, and the Beta-distribution parameterization for bounded steering and acceleration is technically appropriate. Numerically, the weighted BC agent attains a $100\%$ driving score on the training town and a dramatic $98\%$ score on the unseen town, while standard BC collapses to $34\%$, per Table 1.
Experimental scale and rigor are the biggest weaknesses. Training on just eight routes and roughly 36 minutes of driving invites over fitting and raises doubts about statistical reliability; the reported test-town mean of $98 \pm 0.03$ with virtually no variance is implausibly clean for a novel environment. The framing of the prior-art gap is inaccurate: while the authors correctly note that their own earlier work ignored traffic lights, Codevilla et al. explicitly benchmark traffic-light scenarios in CARLA, and Chauffeurnet's mid-level representation includes traffic-light states. The paper also lacks any discussion of how the KDE bandwidth was chosen, how the loss was calibrated, or how many random seeds produced the final table, severely limiting scientific interpretability.
The evidence supports the narrow claim that inverse-KDE weighting helps on this specific dataset, but the comparative analysis is thin. The authors do not evaluate simpler alternatives such as oversampling rare trajectories, inverse-frequency weighting without KDE, or cost-sensitive learning baselines. Likewise, promising related methods like implicit behavioral cloning are relegated to future work rather than empirical comparison. Because the paper mischaracterizes prior work as ignoring traffic lights entirely, the novelty claim is weaker than the experiments actually demonstrate.
Reproducing the results would be difficult. Although the network architecture is diagrammed in Figures 2 and 3, the paper omits optimizer settings, learning rate, batch size, loss scaling, and total training epochs. The KDE module is described mathematically but lacks bandwidth selection criteria or kernel implementation details, both of which strongly affect the density estimates and the resulting sample weights. A video page is provided, yet no code repository or dataset download is mentioned, forcing an independent researcher to re-implement the expert data collection pipeline from scratch.
This paper investigates the use of Behavior Cloning (BC) for autonomous driving from a bird's-eye view (BEV) perspective in simulated urban environments. BC uses supervised learning to mimic expert driving behaviors. Previous works have applied BC in the CARLA simulator but did not fully address the challenges of traffic light compliance. Our approach enhances BC by integrating a kernel density estimator to adjust training sample weights based on action density, thereby improving the learning of rare but critical actions such as stopping at red lights and accelerating at green lights, specially in scenarios of scarce number of expert demonstrations. Using BEV inputs, which provide an abstract top-down view of the driving environment, our method simplifies the policy learning process. The trained convolutional neural network (CNN) outputs steering and acceleration actions based on these BEV inputs and additional state variables. Experimental results in the CARLA simulator demonstrate that our weighted BC method significantly improves driving performance, achieving higher route completion compared to standard BC. This weighted approach proved to be crucial in learning correct driving behaviors, particularly in test environments not encountered during training, highlighting its potential for enhancing autonomous vehicle navigation.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.