NVIDIA-backed RL framework teaches autonomous cars to drift
New hierarchical RL method cuts lap times by optimizing drift control for autonomous rally cars.
A new paper by Sheng Zhao, Bolin Zhao, Xiaodong Wu, and Chen Lv proposes a planning-control framework for autonomous vehicle drifting with minimum-lap-time performance. The approach generates an optimal drift trajectory and uses it as training data for a deep reinforcement learning controller. Because drifting involves extreme sideslip angles, the method uses a track-guided, progressive training process that starts with a drift control policy, moves to a drift corner policy, and finally builds a full drift race policy. The reward combines instant and end rewards tied to the minimum-lap-time objective. Simulation results show the agent learns a drift racing policy that maintains vehicle control while effectively reducing lap time.
- Hierarchical RL framework (TgRL) enables autonomous vehicles to perform rally drifting while optimizing lap times
- Three-stage training: drift control → corner policy → full race policy with minimum-lap-time objectives
- Simulation results show improved lap times while maintaining vehicle stability in loose-surface conditions
Why It Matters
This breakthrough could accelerate autonomous rally racing development and improve emergency maneuvering systems in production vehicles.