티스토리 뷰

djjin/Mujoco

Let's do MuJoCo

djjin 2021. 12. 3. 13:56

21년 10월 18일 딥마인드는 자사 블로그를 통해 Mujoco를 인수했다고 발표했다.

 

Opening up a physics simulator for robotics

As part of DeepMind's mission of advancing science, we have acquired the MuJoCo physics simulator and are making it freely available for everyone, to support research everywhere.

deepmind.com

 

Mujoco(Multi-Joint Dynamics-with-Contact)는 물리적 현상을 자연스럽게 나타내는 시뮬레이터로 유명하며, 주요 엔진이 C 언어로 작성되어 다른 아키텍처와 연동이 쉬운 편이다. 무료 시뮬레이터가 되었기에 Mujoco를 활용한 연구들이 활발하게 진행될 것으로 예상된다.

 

필자 또한 Mujoco 안에서 매니퓰레이터 모션 제어를 연구하고 있고, 올해 7월부터 시작했기에 발만 담근 수준이지만 이제 시작하려는 분들을 위해 지금까지 삽질해온 것들, 공부한 것들을 정리하고자 한다.

최종 목표는 Mujoco에서 Task and Motion Planning 분야를 연구할 수 있는 환경을 만드는 것이다.

 

코드는 Python으로 작성될 것이고, OpenAI에서 mujoco-py라는 Python wrapper 패키지를 만들어 주셔서 Python으로 작성할 수가 있다. (감사합니다 😀 )

 

GitHub - openai/mujoco-py: MuJoCo is a physics engine for detailed, efficient rigid body simulations with contacts. mujoco-py al

MuJoCo is a physics engine for detailed, efficient rigid body simulations with contacts. mujoco-py allows using MuJoCo from Python 3. - GitHub - openai/mujoco-py: MuJoCo is a physics engine for det...

github.com

 

앞으로의 계획을 미리 귀띔하자면 

  • Mujoco, mujoco-py 설치
  • Mujoco file 설명
  • Kinematics
  • Manipulator (Franka panda, Kuka, UR5e, Sawyer, etc..) 불러오기
  • Collision Check
  • Joint Position Control
  • Cartesian Planning
  • RRT* Motion Planning
  • Pick and Place 모듈 개발하기
  • Task and Motion Planning 데모

Manipulation을 하기 위한 기본적인 내용에 대해 지금까지 공부한 것들을 토대로 정리하고자 한다.

우선 기본적인 로봇 공학 지식을 가지고 있어야 다음의 연구(강화학습을 적용한플래닝 알고리즘 연구 등)를 할 수 있다고 생각한다.

 

다시 한번 말하면, 최종 목표는 Mujoco 환경에서 Manipulation 제어를 하는 것이고 쓸모 있는 로봇을 개발하기 위해 노력할 것이다. 부족한 점이 많겠지만, 이제 시작하려는 분들에게는 도움이 될 것이라 생각한다.

댓글
공지사항
최근에 올라온 글
최근에 달린 댓글
Total
Today
Yesterday
링크
«   2024/05   »
1 2 3 4
5 6 7 8 9 10 11
12 13 14 15 16 17 18
19 20 21 22 23 24 25
26 27 28 29 30 31