Skip to content

Solving long-horizon tasks in the Franka Kitchen gym environment via Imitation and Reinforcement Learning

License

Notifications You must be signed in to change notification settings

mobinajamali/franka-HRL

Repository files navigation

franka-HRL

Solving Multi-Stage, Long-Horizon Robotic Tasks in the Franka Kitchen Gym Environment

This repository implements a hybrid Imitation Learning (IL) and Reinforcement Learning (RL) approach to address complex, multi-step robotic tasks in the sparse-reward environment of the Franka Kitchen. It integrates human experiences and hierarchical learning to enhance long-horizon task performance.

Franka-HRL


Key Features

🛠 Three-Phase Learning Approach

  1. Imitation Learning Phase:

    • Creating goal-conditioned hierarchical policies based on human demonstrations.
    • Using a game controller interface to collect human-piloted robot actions for tasks like opening the microwave, building a comprehensive replay buffer.
    • Weighted replay buffers prioritize human-generated experiences during early training stages.
  2. Reinforcement Learning Phase:

    • Fine-tuning policies with the Soft Actor-Critic (SAC) algorithm for efficient long-horizon task performance.
    • Dynamic reliance on human data ensures smooth transitions from imitation to independent policy learning.
  3. Hierarchical Reinforcement Learning:

    • Implementing a meta-agent to manage a static list of task-specific policies.
    • Coordinating and loading the appropriate sub-policy to solve individual subtasks in a multi-task environment.

Inspired By:

This project is inspired by the paper:
Relay Policy Learning: Solving Long-Horizon Tasks via Imitation and Reinforcement Learning

About

Solving long-horizon tasks in the Franka Kitchen gym environment via Imitation and Reinforcement Learning

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages