Bandwidth Allocation for Cloud-Augmented Autonomous Driving

TL;DR

The paper tackles bandwidth-constrained cloud augmentation for autonomous driving by introducing TURBO, a system that jointly optimizes cloud-model selection and per-task data transmission to maximize driving utility under finite network bandwidth. It defines model- and service-level utility curves, and formulates an ILP to allocate bandwidth and pick cloud configurations in real time, allowing multiple cloud options and data compression. Empirical results on the Waymo Open Dataset show TURBO achieving up to 15.6 percentage-point accuracy gains over on-car-only baselines and gains under realistic 5G-like RTTs, while dynamic utility policies offer additional improvements under varying environments. The work highlights practical considerations for cloud-assisted AVs, including latency, data transfer costs, and reliability, and establishes a general approach that could extend to other data-intensive, networked AI settings.

Abstract

Autonomous vehicle (AV) control systems increasingly rely on ML models for tasks such as perception and planning. Current practice is to run these models on the car's local hardware due to real-time latency constraints and reliability concerns, which limits model size and thus accuracy. Prior work has observed that we could augment current systems by running larger models in the cloud, relying on faster cloud runtimes to offset the cellular network latency. However, prior work does not account for an important practical constraint: limited cellular bandwidth. We show that, for typical bandwidth levels, proposed techniques for cloud-augmented AV models take too long to transfer data, thus mostly falling back to the on-car models and resulting in no accuracy improvement. In this work, we show that realizing cloud-augmented AV models requires intelligent use of this scarce bandwidth, i.e. carefully allocating bandwidth across tasks and providing multiple data compression and model options. We formulate this as a resource allocation problem to maximize car utility, and present our system \sysname which achieves an increase in average model accuracy by up to 15 percentage points on driving scenarios from the Waymo Open Dataset.

Figures (11)

• Figure 1: TURBO offers larger and more accurate cloud models to the on-car control system, optimally selecting the subset of inputs to send and models to run in the cloud and apportioning bandwidth.
• Figure 2: Utility curves for a service with 3 obj. detection models.
• Figure 3: We consider 35 model configurations with different tradeoffs in runtime, accuracy, and data transfer size: ED1 runs on-vehicle while ED2, ED4, ED6, and ED7x execute in the cloud. We generate additional configurations by applying lossless PNG and lossy JPEG compression to either the original image or the pre-processed model inputs. Our JPEG quality factors are 95, 90, 75, and 50. We find that compression significantly reduces the amount of data to transfer at the cost of increased runtime and reduced accuracy.
• Figure 4: Average accuracy across services for TURBO compared to baselines of varying naivete as bandwidth increases. We assume an RTT of 20ms, object detection SLOs of 150ms, and motion planning SLO of 250ms, averaging performance across all scenarios.
• Figure 5: Incremental benefit from TURBO's design decisions.