A Remote Control Strategy for an Autonomous Vehicle with Slow Sensor Using Kalman Filtering and Dual-Rate Control

Ángel Cuenca, Wei Zhan, Julián Salt, José Alcaina, Chen Tang, Masayoshi Tomizuka

TL;DR

This paper presents a novel remote control system for autonomous vehicles (AVs) that tackles challenges like slow sensors, network delays, and packet disorder. By integrating an Extended Kalman Filter (EKF) with an h-step-ahead prediction, a dual-rate control strategy, and a pure pursuit path-tracking algorithm, the system achieves nominal control performance while significantly reducing network bandwidth usage.

What Problem Does This Paper Address?

Controlling autonomous vehicles (AVs) remotely, especially for tasks like path tracking, presents a significant engineering challenge. The core issue is bridging the gap between a local AV system and a remote control server using a communication network. This network introduces several problems:

  • Time-varying delays: The time it takes for data to travel between the AV and the server can fluctuate unpredictably.
  • Packet disorder: Data packets might arrive out of order.
  • Packet loss: Packets can be lost entirely.
  • Network bandwidth constraints: The amount of data that can be transmitted is limited.
  • Slow sensors: Some critical sensors, like cameras used for positioning, provide data at a much slower rate than required for real-time control.

These issues can degrade control performance, leading to inaccurate path tracking and instability. Traditional control systems often struggle to maintain desired performance under such conditions. This paper aims to develop a robust remote control solution that can overcome these network-induced limitations and sensor speed constraints to achieve high-quality AV path tracking.

Methods

The proposed solution is a sophisticated integration of several key techniques:

  • Extended Kalman Filter (EKF) with h-step-ahead prediction: The EKF is used to fuse data from various sensors (encoders, IMU, camera) to estimate the AV's state (position, orientation, velocities) while reducing noise. Crucially, it incorporates an "h-step-ahead" prediction stage. This allows the system to predict future states, which helps in compensating for network delays by enabling control actions to be computed based on anticipated future states rather than just current, potentially delayed, information.
  • Dual-Rate Control: This technique addresses the issue of slow sensors. It decouples the control system into two different time periods: a faster actuation period (T) for the motors and a slower sensing period (NT). This allows the AV to react quickly to control commands even if its position is updated less frequently. The dual-rate approach also helps in reducing network traffic by transmitting data less often.
  • Packet-Based Control: This strategy involves sending data in packets, often grouping multiple pieces of information. In this work, it's used to send a set of future path references from the remote server to the AV and a set of future state estimations from the AV back to the server. This, combined with the h-step-ahead prediction, helps manage network delays.
  • Pure Pursuit Path Tracking Algorithm: This is a well-established algorithm used to guide the AV along a desired path. It calculates the necessary steering commands based on the AV's current position and a target point ahead on the path.

The overall system architecture splits tasks between a local side (on the AV) and a remote server. The remote server handles higher-level tasks like path generation, while the local side manages state estimation, path tracking, and control, all while dealing with network imperfections.

Datasets

Not discussed in the paper. The authors use a simulated Lego robot model for their experiments.

Key Insights and Ideas

  • Predictive control is key for network delays: The h-step-ahead prediction stage in the EKF is crucial. It allows the controller to "look into the future" and make decisions based on predicted states, effectively masking the impact of time-varying network delays.
  • Dual-rate control is a practical solution for slow sensors: By separating sensing and actuation rates, the system can maintain good control performance even when a critical sensor (like a camera) is slow, without sacrificing responsiveness.
  • Integration is more than the sum of its parts: The paper demonstrates that combining EKF, prediction, dual-rate control, and packet-based strategies creates a synergistic effect, allowing the system to achieve robust performance under challenging conditions where individual techniques might fail.
  • Bandwidth savings are significant: By using a slower sensing period (NT) and transmitting data less frequently, the system achieves a substantial reduction in network bandwidth usage (up to 50% in their simulations), which is vital for practical wireless applications.
  • Robustness to noise and delays is achievable: The proposed integrated approach shows remarkable resilience to both process/measurement noise and network-induced delays, maintaining near-nominal performance.

Key Results

The paper presents simulation results comparing their proposed dual-rate control with EKF, prediction, and packet-based strategies against simpler single-rate and dual-rate controllers under various conditions.

  • Nominal Performance: A single-rate control at a fast rate (T=0.1s) without delays or noise served as the benchmark for nominal performance.
  • Single-Rate Degradation: When the sampling period was increased to NT=0.2s (slower sensing), single-rate control performance significantly worsened.
  • Dual-Rate Improvement: The dual-rate control (NT=0.2s, T=0.1s) recovered much of the lost performance, achieving results similar to the nominal case.
  • Impact of Delays: Introducing time-varying delays to the dual-rate control caused performance degradation.
  • Proposed Solution's Success: The full proposed solution (dual-rate control + EKF + prediction + packet-based control) with delays and a lower level of noise demonstrated significantly improved performance, closely matching the nominal behavior.
  • Bandwidth Savings: The dual-rate approach (sampling twice slower) reduced network usage by 50%.
  • Cost Indexes: Quantitative analysis using cost indexes (J1, J2, J3) showed that the proposed method (scenario 'e') achieved J1 and J2 values comparable to the ideal nominal case ('a'), while J3 (time to destination) was similar across most scenarios. Scenarios 'b' and 'd' (single-rate NT and dual-rate with delays) showed the worst performance in J1 and J2.

Limitations and Future Work

  • Noise Sensitivity: The paper notes that at a higher level of noise (F=0.45), the robustness of the control proposal could be compromised. This scenario was not considered for further analysis.
  • Model Accuracy: The effectiveness of the EKF and the control system relies on an accurate system model. Deviations from the assumed model could impact performance.
  • Future Work: The paper doesn't explicitly detail future work directions, but the demonstrated robustness and efficiency suggest potential applications in more complex AV scenarios, integration with advanced AI for decision-making, and real-world implementation on physical platforms. The authors also mention that the control signal was affected by noise at higher levels, implying that further research into noise reduction or more robust control strategies could be beneficial.

Summary updated

BibTeX
@inproceedings{cuenca2019remote,
  author = {Ángel Cuenca and Wei Zhan and Julián Salt and José Alcaina and Chen Tang and Masayoshi Tomizuka},
  title = {{A Remote Control Strategy for an Autonomous Vehicle with Slow Sensor Using Kalman Filtering and Dual-Rate Control}},
  booktitle = {Sensors},
  year = {2019},
  url = {https://openreview.net/forum?id=PMGq0l2acPs},
}
Figure