A Curriculum Approach to Bridge the Reality Gap in Autonomous Driving Decision-Making based on Deep Reinforcement Learning

1Universidad de Alcalá

Abstract

Decision-Making is a fundamental topic in the domain of Autonomous Driving where significant challenges must be tackled due to the variable behaviours of surrounding agents and the wide array of encountered scenarios. The primary aim of this work is to develop a hybrid Decision-Making architecture able to be validated on a real vehicle that marries the reliability of classical techniques with the adaptability of Deep Reinforcement Learning approaches. To address the crucial transition from simulated training environments to real-world applications, this research employs a Curriculum Learning approach, facilitated by the deployment of Digital Twins and Parallel Intelligence technologies, significantly narrowing the Reality Gap and enhancing the applicability of learned behaviours. The viability of this approach is evidenced through a Parallel Execution, wherein simulated and real-world tests are conducted simultaneously. Specifically, our approach consistently surpasses the performance benchmarks set by existing frameworks in the literature within SMARTS, achieving success rates over 91\%. Additionally, it completes various scenarios in CARLA up to 50\% faster than the Autopilot, demonstrating improved comfort and safety.

Experiments in SUMO

Lightweigh simulation for higl-level behaviours learning.

Unprotected Left Turn Scenario

The green agent must merge into the left lane.

Three Lane Merge Scenario

The green agent must merge into the traffic.

Three Lane Road Scenario

The green agent must reach the end of the road.

Roundabout Scenario

The green agent must merge into the roundabout and leave it in the last exit.

SUMO Quantitative Results

The results in terms of success rate (%) and average completion time (sec). An ablation study between different DRL agents and a comparison to representative SOTA proposals are presented. The training pregression of the rewards is also represented.

Experiments in CARLA

Realistic simulation including vehicle dynamics

Town 03 Crossroad Scenario

The ego vehicle must cross the intersection while vehicles are coming from both sides.

The ego vehicle stops due to the adversarial vehicles and starts moving when it identifies a gap. The control signals and comfor metrics are represented bellow.

Town 03 Merge Scenario

The ego vehicle must merge into the right lane while vehicles are coming from left side.

The ego vehicle stops due to the adversarial vehicles and starts moving when it identifies a gap. The control signals and comfor metrics are represented bellow.

Town 04 Lane Change Scenario

The ego vehicle must drive until the end of the road while vehicles are driving in this road.

The ego vehicle changes lane two times resulting in the shown control signals.

Town 03 Roundabout Scenario

The ego vehicle must merge into the roundabout and leave it in the second exit.

The control signals are repreente of the vehicle approaching and mergin the roundabout.

Town 03 Concatenated Scenario

The ego vehicle drives through all the previously introduced scenarios in a concatenated way.

The fisrt charts shows the ego vehicle performance while the second one presentd the CARLA Autopilor signals under the same scenario.

CARLA Quantitative Results

The comparison using different metrics betwenn our proposal and the CARLA Autopilot.

Parallel Execution Experiments

Validation of our proposal with a Parallel Execution in a merge scenario in our University Campus using our Real Vehicle.

Low Traffic Flow

Front View

Side View

Control Graphics

Mixed Traffic Flow

Front View

Side View

Control Graphics

BibTeX

@article{gutierrez2024curricular,
  author    = {Gutiérrez-Moreno, Rodrigo and Barea, Rafael and López-Guillén, Elena and Arango, Felipe and Bergasa, Luis Miguel},
  title     = {A Curriculum Approach to Bridge the Reality Gap in Autonomous Driving Decision-Making based on Deep Reinforcement Learning},
  journal   = {IEEE Transactions on Intelligent Vehicles (T-IV)},
  year      = {2024},
}