Skip to content

Latest commit

 

History

History
53 lines (34 loc) · 1.67 KB

README.md

File metadata and controls

53 lines (34 loc) · 1.67 KB
Raw

Trajectory Planning using HER & Reward Engineering

Trajectory planning based on Reinforcement Learning with Hindsight Experience Replay, Prioritized Experience Replay & Dense Reward Engineering to solve openai-gym robotics "FetchReach-v1" environment using PyTorch & Tensorflow2.

Reinforcement Learning Algorithms

  1. Dense Reward Engineering: Engineered vector based distance measure to replace sparse rewards.

  2. Hindsight Experience Relay (HER): Implemented HER Future Strategy based goal sampling for buffer augmentation.

  3. Prioritized Experience Relay (PER): Samples and optimizes the past experiences ended with errors to get better future rewards.

Agent Profiles

  1. Dense Reward Engineering

    DDPG Agent
    PER + DDPG Agent

  2. Hindsight Experience Repay

    DDPG Agent
    PER + DDPG Agent

Play Preview

  • Previews from older implementation in TF2.

    Dense Rewards HER

Dependencies

Install dependencies using:

pip3 install -r requirements.txt
  1. Additionally install 'mujoco_py' according to 'https://github.com/openai/mujoco-py'

Developer