Enhance Deep Reinforcement Learning with Denoising Autoencoder for Self-Driving Mobile Robot

Abstrak

Over the past years, self-driving mobile robots have captured the interest of researchers, prompting exploration into their multifaceted implementation. They have the potential to revolutionize transportation by mitigating human error and reducing traffic accidents. The process of deploying self-driving mobile robots can be divided into several steps, such as algorithm design, simulation, and real-world application. This research paper presents a simulation using DonkeyCar on the Mini Monaco track, employing a Soft Actor-Critic (SAC) alongside a denoising autoencoder. At this point, it is limited to the simulation, serving as a proof of concept for further research with hardware implementation. The simulation verifies that relying solely on SAC for the convergence of policy is not sufficient; it yields a mean episode length of only 28.82 steps and a mean episode reward of 0.7815. The simulation ended after 3557 steps due to the inability of SAC alone to converge, without completing a single lap. Later, by integrating the denoising autoencoder, convergence of policy can be achieved. It enables DonkeyCar to adeptly track the lane of the circuit. The denoising autoencoder plays an important role in accelerating the convergence of transfer learning. Notably, the mean reward per episode reached 2380.4387, with an average episode length of 771.71 and a total of 114357 steps taken. DonkeyCar manages to complete several laps. These results affirm the effectiveness of SAC with a denoising autoencoder in enhancing the performance of self-driving mobile robots.

Kata kunci

• Self-Driving Mobile Robot
• Deep Reinforcement Learning
• Donkeycar Simulation
• Soft Actor-Critic
• Denoising Autoencoder.