Computational Optimization of Image-Based Reinforcement Learning for Robotics.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Science

Background:

• Deep learning significantly impacts robotics, with a trend towards large, pretrained models.
• Existing large models are incompatible with robotic hardware and impede low-latency control.
• Computational intensity of deep learning models hinders real-time robotic applications.

Purpose of the Study:

• To enhance the computational efficiency of deep learning models in reinforcement learning (RL) for robotics.
• To enable the deployment of deep learning models on resource-constrained robotic systems.
• To demonstrate real-world closed-loop control using optimized models on edge devices.

Main Methods:

• Reduced model architecture complexity by decreasing layers and altering structure.
• Downscaled input resolution to decrease computational load.
• Applied weight quantization, comparing post-training quantization and quantization-aware training.

Main Results:

• Optimization strategies improved computational efficiency by approximately 300 times.
• Functional performance of the system was slightly improved post-optimization.
• Demonstrated closed-loop control on a real robot using a Jetson Xavier NX edge device.

Conclusions:

• Computational efficiency of deep learning models in robotics can be significantly improved.
• Optimized models are suitable for real-time control tasks on edge computing platforms.
• The proposed strategies balance computational cost and functional performance in robotic RL.