Consecutive Inertia Drift of Autonomous RC Car via Primitive-based Planning and Data-driven Control

Yiwen Lu, Bo Yang, Jiayun Li, Yihan Zhou, Hongshuai Chen, Yilin Mo

Abstract

Achieving aggressive maneuverability in autonomous vehicles requires navigating the limits of vehicle dynamics. This paper presents a novel framework for "inertia drift"—the rapid transition between counter-clockwise and clockwise drifting states. By combining a computationally efficient primitive-based planner with a data-driven controller that interpolates real-world trajectories, the proposed system overcomes the challenges of high nonlinearity and modeling inaccuracies. The framework was successfully validated on a 1/10-scale RC car, demonstrating stable, continuous maneuvers across complex paths such as 8-shaped and "Olympic Rings" configurations in both simulation and physical environments.

Introduction

Drifting represents the extreme edge of vehicle dynamics, where a car is intentionally pushed beyond its traction limits to maximize maneuverability. While autonomous systems have previously mastered sustained drift (maintaining a circular skid) and transient drift (single maneuvers like drift parking), "inertia drift" remains a significant hurdle. Inertia drift involves the aggressive transition from one drift direction to another—for example, snapping from a left-hand slide into a right-hand slide.

This maneuver is exceptionally difficult to automate for two primary reasons. First, the vehicle must traverse a series of highly nonlinear, unstable transient states rather than settling into a predictable equilibrium. Second, standard linearized models often fail to capture the physics of these extreme regimes, leading to a "reality gap" where simulation-based controllers fail on physical hardware. This research aims to bridge this gap, enabling autonomous agents to navigate multi-turn paths without losing speed or stability.

Methodology

The researchers propose a two-layered architecture consisting of a high-level primitive-based planner and a low-level data-driven controller.

Primitive-based Planning

Rather than attempting to solve complex trajectory optimization problems in real-time, the system utilizes a pre-computed library of "motion primitives."

  • Offline Generation: Primitives are initially generated using trajectory optimization (iLQR) based on a bicycle model. These are then "calibrated" by executing them on physical hardware to record feasible, real-world trajectories.
  • Online Concatenation: A Finite State Machine (FSM) monitors the vehicle's state in real-time. As the car approaches a transition point, the planner selects the most appropriate primitive from the library based on geometric criteria. This approach allows for high-frequency onboard planning at 100Hz.

Data-driven Control

To address the mismatch between theoretical paths and real-world physics, the system employs a data-driven interpolation strategy:

  1. Weighted Averaging: The controller identifies several candidate primitives from the library that match the current state. It then calculates a weighted average of these trajectories to generate a near-feasible reference path.
  2. Soft Dynamic Time Warping (DTW): To ensure smooth transitions between trajectories of varying lengths and speeds, the authors utilize a differentiable version of DTW to align and average the primitives.
  3. Tracking: A feedforward/feedback controller tracks the interpolated reference, making real-time adjustments to steering angles and wheel speeds based on the observed error.

Experimental Setup

The framework was tested using a custom-built Primitive Library. In simulation, a high-fidelity four-wheel model with Pacejka tire dynamics was used. For physical validation, the researchers utilized a 1/10-scale four-wheel-drive RC car powered by an NVIDIA Jetson TX2 and a VESC motor controller. The library contains 54 unique inertia drift primitives for each platform, covering a wide range of initial speeds and sideslip angles.

Key Results and Insights

The study yielded several critical insights into the nature of aggressive autonomous driving:

  • Stability and Robustness: The proposed method maintained bounded trajectories over multiple laps of an 8-shaped path. In comparative tests, state-of-the-art nonlinear model inversion and reinforcement learning (PPO) methods frequently diverged or failed to maintain the drift during transitions.
  • Versatility: The system successfully navigated complex layouts, including three-circle paths and the "Olympic Rings" configuration, demonstrating that complex maneuvers can be decomposed into a "Lego-set" of pre-validated segments.
  • Bridging the Reality Gap: By interpolating between real trajectories that have already been proven on hardware, the authors bypassed the need for perfect mathematical system identification, resulting in a controller that is inherently more resilient to modeling errors.

Conclusion and Future Work

This research demonstrates that combining primitive-based planning with data-driven interpolation allows autonomous vehicles to perform complex, high-speed maneuvers that were previously the domain of expert human drivers. While the current system is stable and aggressive, future work will focus on optimizing these paths for the "perfect racing line" to minimize lap times. Additionally, the researchers plan to explore how these primitives can be adapted to varying surface friction and integrated into competitive, multi-agent racing scenarios.

Summary updated

BibTeX
@inproceedings{lu2023consecutive,
  author = {Yiwen Lu and Bo Yang and Jiayun Li and Yihan Zhou and Hongshuai Chen and Yilin Mo},
  title = {{Consecutive Inertia Drift of Autonomous RC Car via Primitive-based Planning and Data-driven Control}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10341461},
}
Figure