Enhancing the Performance of Multi-Vehicle Navigation in Unstructured Environments using Hard Sample Mining

TL;DR

This work tackles multi-vehicle navigation in unstructured environments without traffic signals by leveraging a Graph Neural Network (GNN) to model vehicle interactions and an optimization-based labeling pipeline. A warm-start strategy pre-trains on simple scenarios and uses hard-sample mining to identify the most informative training samples, achieving up to a 10x reduction in labeled data and up to 1.8x faster optimization when labeling hard cases. The approach outperforms baselines across diverse vehicle counts, showing improved success-to-goal rates and lower collision rates, with added robustness to steering-noise. The method offers a data-efficient path to deploying autonomous multi-vehicle control in non-road contexts such as warehouses or underdeveloped road networks.

Abstract

Contemporary research in autonomous driving has demonstrated tremendous potential in emulating the traits of human driving. However, they primarily cater to areas with well built road infrastructure and appropriate traffic management systems. Therefore, in the absence of traffic signals or in unstructured environments, these self-driving algorithms are expected to fail. This paper proposes a strategy for autonomously navigating multiple vehicles in close proximity to their desired destinations without traffic rules in unstructured environments. Graphical Neural Networks (GNNs) have demonstrated good utility for this task of multi-vehicle control. Among the different alternatives of training GNNs, supervised methods have proven to be most data-efficient, albeit require ground truth labels. However, these labels may not always be available, particularly in unstructured environments without traffic regulations. Therefore, a tedious optimization process may be required to determine them while ensuring that the vehicles reach their desired destination and do not collide with each other or any obstacles. Therefore, in order to expedite the training process, it is essential to reduce the optimization time and select only those samples for labeling that add most value to the training. In this paper, we propose a warm start method that first uses a pre-trained model trained on a simpler subset of data. Inference is then done on more complicated scenarios, to determine the hard samples wherein the model faces the greatest predicament. This is measured by the difficulty vehicles encounter in reaching their desired destination without collision. Experimental results demonstrate that mining for hard samples in this manner reduces the requirement for supervised training data by 10 fold. Videos and code can be found here: \url{https://yininghase.github.io/multiagent-collision-mining/}.

Figures (4)

• Figure 1: Shows the overview of our framework. Step 1: A pre-trained model trained on a simpler subset of data is first executed on multiple complicated scenarios to infer the corresponding trajectories. Scenarios producing no collision in their trajectories are depicted by a green border, while others are shown in red. The points of collision are depicted by red dots. For simplicity, we have only shown two out of the many scenarios. Step 2: Samples with no or minimal collisions are filtered out, leaving only the examples that were difficult for the pre-trained model to predict the correct control labels. Step 3: These remaining difficult/hard samples are passed to an optimization algorithm to produce the corresponding control labels (Shown in dotted red color). The dotted blue lines are the predictions from the pre-trained model that are used for warm-starting the iterative optimization process. Step 4: This newly labeled additional data is used to train a Graphical Neural Network (GNN) model, to improve performance. Step 5: Inference is done on unseen samples that can have even more vehicles than in the training set. Note that the black arrows depict the flow of the training process, while the green arrow shows the inference process. Out of the entire pipeline, only Step 5 is needed at test time.
• Figure 2: Shows a qualitative visualization of the trajectory traversed by the 3 models for a sample scenario. The red dots on the trajectories indicate collision between the vehicles. Only our model does not have collisions, while the other two have plenty. Corresponding videos can be visualized on the project page: https://yininghase.github.io/multiagent-collision-mining/#Comparison-with-Other-Models
• Figure 3: Shows the normalized probability density function of trajectories at different collision rates for 4 vehicle/obstacle configurations. It can be observed that the distribution of the baseline method trained with additional random data (red curve) is always to the right of the distribution of our method (blue curve). This is more pronounced as the number of vehicles in the scene are increased. This shows that the method trained with random samples has more occurrences with higher collision rates.
• Figure 4: Shows the relative change in success-to-goal ratio when the variance of noise intensity controlled by $\alpha$ is increased from 0 to 0.3. All values are normalized in reference to the results for $\alpha$ = 0. The curves show the mean performance across all the vehicle/obstacle configurations reported in Table \ref{['table: statistic of absolute decreased performance on steering angle noise']} for the baseline and our model trained on hard sampled data. The figure also depicts the standard deviation for each model, as shown by the shaded regions. The results in this figure show that our model is not only relatively more robust to noise in the steering angle as depicted by a slower drop in success-to-goal ratio but also is more stable in its prediction across the different configurations as demonstrated by a lower standard deviation.