AgentGPT: Remote Env Integrated Cloud RL Training

W&B Humanoid-v5 Benchmark (via Internet): Weights & Biases Dashboard

Overview

AgentGPT is a one-click, cloud-based platform for distributed reinforcement learning. It lets you easily host your environment simulators—either locally or in the cloud—and connect them to a central training job on AWS SageMaker. This enables efficient data collection and scalable multi-agent training using a GPT-based RL policy.

Installation

pip install agent-gpt-aws --upgrade

Simulation

• Launch your environment simulator (e.g., Gym, Unity, Unreal) before training begins:
  With this command, your local machine automatically connects to our AgentGPT WebSocket server on the cloud. This real-time connection enables seamless data communication between your environment's state and the cloud training actions, ensuring that everything is ready for the next agent-gpt train command.

   agent-gpt simulate
  

Training & Inference

• Train a gpt model on AWS:

  agent-gpt train
  

• Run agent gpt on AWS:

  agent-gpt infer
  

Configuration

• Config hyperparams & SageMaker:

  agent-gpt config --batch_size 256
  agent-gpt config --role_arn arn:aws:iam::123456789012:role/AgentGPTSageMakerRole
  

• List & Clear current configuration:

  agent-gpt list
  agent-gpt clear
  

Key Features

• Cloud & Local Hosting: Quickly deploy environments (Gym/Unity) with a single command.
• Parallel Training: Connect multiple simulators to one AWS SageMaker trainer.
• Real-Time Inference: Serve a GPT-based RL policy for instant decision-making.
• Cost-Optimized: Minimize expenses by centralizing training while keeping simulations local if needed.
• Scalable GPT Support: Train Actor (policy) and Critic (value) GPT models together using reverse transitions.