Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Agile Autonomous Driving using End-to-End Deep Imitation Learning

Yunpeng Pan, Ching-An Cheng, Kamil Saigol, Keuntaek Lee, Xinyan Yan, Evangelos Theodorou, Byron Boots

TL;DR

This work tackles agile, high-speed off-road autonomous driving using only low-cost sensors by learning an end-to-end DNN policy that imitates a model-predictive controller (MPC). It analyzes online versus batch imitation learning, showing that online IL (via DAgger and a mixed-expert data collection) yields better generalization and robustness to covariate shift, enabling safe, high-speed operation on a dirt track. The autonomous driving system combines a CNN for monocular vision with wheel-speed inputs, trained to map observations directly to steering and throttle without state estimation or online planning, and is validated on a 1/5-scale AutoRally platform. The MPC expert relies on a Sparse Spectrum Gaussian Process dynamics model and Differential Dynamic Programming, providing high-quality demonstrations that guide the learner. Overall, the approach demonstrates data-efficient, real-world autonomous driving with low-cost sensors and highlights the importance of online IL for reliability in stochastic, real-world environments.

Abstract

We present an end-to-end imitation learning system for agile, off-road autonomous driving using only low-cost sensors. By imitating a model predictive controller equipped with advanced sensors, we train a deep neural network control policy to map raw, high-dimensional observations to continuous steering and throttle commands. Compared with recent approaches to similar tasks, our method requires neither state estimation nor on-the-fly planning to navigate the vehicle. Our approach relies on, and experimentally validates, recent imitation learning theory. Empirically, we show that policies trained with online imitation learning overcome well-known challenges related to covariate shift and generalize better than policies trained with batch imitation learning. Built on these insights, our autonomous driving system demonstrates successful high-speed off-road driving, matching the state-of-the-art performance.

Agile Autonomous Driving using End-to-End Deep Imitation Learning

TL;DR

This work tackles agile, high-speed off-road autonomous driving using only low-cost sensors by learning an end-to-end DNN policy that imitates a model-predictive controller (MPC). It analyzes online versus batch imitation learning, showing that online IL (via DAgger and a mixed-expert data collection) yields better generalization and robustness to covariate shift, enabling safe, high-speed operation on a dirt track. The autonomous driving system combines a CNN for monocular vision with wheel-speed inputs, trained to map observations directly to steering and throttle without state estimation or online planning, and is validated on a 1/5-scale AutoRally platform. The MPC expert relies on a Sparse Spectrum Gaussian Process dynamics model and Differential Dynamic Programming, providing high-quality demonstrations that guide the learner. Overall, the approach demonstrates data-efficient, real-world autonomous driving with low-cost sensors and highlights the importance of online IL for reliability in stochastic, real-world environments.

Abstract

We present an end-to-end imitation learning system for agile, off-road autonomous driving using only low-cost sensors. By imitating a model predictive controller equipped with advanced sensors, we train a deep neural network control policy to map raw, high-dimensional observations to continuous steering and throttle commands. Compared with recent approaches to similar tasks, our method requires neither state estimation nor on-the-fly planning to navigate the vehicle. Our approach relies on, and experimentally validates, recent imitation learning theory. Empirically, we show that policies trained with online imitation learning overcome well-known challenges related to covariate shift and generalize better than policies trained with batch imitation learning. Built on these insights, our autonomous driving system demonstrates successful high-speed off-road driving, matching the state-of-the-art performance.

Paper Structure

This paper contains 27 sections, 1 theorem, 26 equations, 10 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Imitation Learning for Autonomous Driving
  4. Problem Definition
  5. Imitation Learning
  6. Online Imitation Learning
  7. Batch Imitation Learning
  8. Comparison of Imitation Learning Algorithms
  9. The Autonomous Driving System
  10. An Algorithmic Expert: Model-Predictive Control
  11. Learning a DNN Control Policy
  12. The Autonomous Driving Platform
  13. Experimental Setup
  14. High-speed Driving Task
  15. Test Track
  16. ...and 12 more sections

Key Result

Lemma 1

Define $d_\pi(s,t) = \frac{1}{T} d_{\pi}^t(s)$ as a generalized stationary time-state distribution, where $d_{\pi}^t$ is the distribution of state at time $t$ when running policy $\pi$. Let $\pi$ and $\pi'$ be two policies. Then where $A_{\pi'}^t(s, a) = Q_{\pi'}^t(s,a) - V_{\pi'}^t(s)$ is the (dis)advantage function at time $t$ with respect to running $\pi'$.

Figures (10)

  • Figure 1: The high-speed off-road driving task.
  • Figure 2: System diagram.
  • Figure 3: The DNN control policy.
  • Figure 4: The AutoRally car and the test track.
  • Figure 5: Examples of vehicle trajectories, where online IL avoids the crashing case encountered by batch IL. (b) and (c) depict the test runs after training on 9,000 samples.
  • ...and 5 more figures

Theorems & Definitions (2)

  • Lemma 1
  • Definition 1