LOKI-G: An Edge AI Training Framework for Physical Machines, Integrating Imitation Learning and Reinforcement Learning

LOKI-G is an edge AI training framework specifically designed for physical machines. It integrates imitation learning and reinforcement learning, and uses continuous-time neural networks and neural circuit policy techniques to solve AI training problems under resource constraints of edge devices. This project is open-source (MIT license), allowing models to learn directly on physical hardware, achieving low latency and high privacy protection, and is suitable for industrial automation, robotics, and other fields.

Background: Challenges in Intelligent Physical Machines

In the fields of industrial automation and robotics technology, deploying artificial intelligence models to physical machines faces unique challenges: traditional deep learning models require large amounts of computing resources and are difficult to run in real-time on edge devices; the physical environment is complex with high safety requirements, so training needs to be efficient and controllable.

To address these issues, the LOKI-G project combines advanced machine learning algorithms with physical hardware requirements to provide a complete solution for edge AI applications.

Core Technologies: Hybrid Learning Strategy

LOKI-G integrates Imitation Learning (IL) and Reinforcement Learning (RL):

• Imitation Learning Phase: Learns from demonstration data, runs for 10 epochs by default;
• Reinforcement Learning Phase: Optimizes strategies through interaction with the environment, runs for 10 epochs by default;
• Intelligent Switching Mechanism: Supports fixed (default at the 18th iteration) or random switching timing.

Neural Network Architecture

Two technologies adapted for physical systems are used:

• Continuous-Time Neural Network: Directly models the dynamic characteristics of physical systems and handles continuous change processes;
• Neural Circuit Policy: Inspired by biological nervous systems, it is efficient in computation and interpretable, suitable for real-time control.

These technologies are provided via the ncps package, and dependencies can be installed via pip.

Practical Applications and Configuration

LOKI-G is compatible with various physical machines. Key configuration points:

• Environment Requirements: Python 3.7+, dependencies installed via requirements.txt;
• Environment Configuration: A custom environment definition file (--env_file parameter) is required;
• Action Output: The number of actions is defined via --num_outputs (default 6);
• Training Visualization: Enable the --render option to observe the environment state in real-time.

Model Saving and Deployment

Trained models are saved in TensorFlow format in the saved_models directory one level above the script's running directory. The standardized format facilitates subsequent deployment and inference, allowing easy integration into production systems.

Technical Significance and Future Outlook

LOKI-G advances the field of edge AI and provides a feasible path for the intelligence of physical machines. In fields such as industrial automation, robotics, and autonomous driving, it solves the problem of training complex AI models in resource-constrained environments.

As edge computing capabilities improve, such technologies will be applied in more scenarios, promoting deep integration between the physical world and AI.

Summary

LOKI-G is an edge AI training framework for physical machines. Through hybrid learning strategies and advanced neural network architectures, it addresses the core challenges of AI training on edge devices. For developers and researchers who want to deploy AI capabilities on physical hardware, it is an open-source project worth paying attention to.