Improve collision-free path planning for robotic arms used for waste sorting by leveraging the CUDA-accelerated parallel planning library cuRobo. On top of that (scope: Master Thesis) we want to train a reinforcement learning policy that performs object sorting based on the state of the tracked items on the conveyor belt, outputting the end-effector poses to be tracked by cuRobo.
The Autonomous River Cleanup (ARC) is a student-led initiative supported by the Robotic Systems Lab that aims to remove waste from rivers. By joining ARC, you’ll work on the Mobile Autonomous Recycling Container (MARC), which uses two robotic arms to sort waste by material (plastic, metal, glass, etc.).
Currently, we perform object tracking, high-level planning (assigning items to the robot arms), and robot control as separate processes. The project aims to combine high-level planning and robot control into an end-to-end reinforcement learning (RL) based policy. Based on the state of the tracked objects and the robot arms, this policy should decide the item assignment, picking order, and the robot end-effector poses of both robots. For the low-level joint control of the robot arms, we plan to leverage the GPU-accelerated path planner cuRobo [1] to find collision-free paths to track the end-effector poses resulting from the RL policy.
[1] Nvidia, cuRobo: CUDA Accelerated Robot Library, https://curobo.org/, last access: 05.09.2024
The Autonomous River Cleanup (ARC) is a student-led initiative supported by the Robotic Systems Lab that aims to remove waste from rivers. By joining ARC, you’ll work on the Mobile Autonomous Recycling Container (MARC), which uses two robotic arms to sort waste by material (plastic, metal, glass, etc.). Currently, we perform object tracking, high-level planning (assigning items to the robot arms), and robot control as separate processes. The project aims to combine high-level planning and robot control into an end-to-end reinforcement learning (RL) based policy. Based on the state of the tracked objects and the robot arms, this policy should decide the item assignment, picking order, and the robot end-effector poses of both robots. For the low-level joint control of the robot arms, we plan to leverage the GPU-accelerated path planner cuRobo [1] to find collision-free paths to track the end-effector poses resulting from the RL policy. [1] Nvidia, cuRobo: CUDA Accelerated Robot Library, https://curobo.org/, last access: 05.09.2024
During your time at ARC, you will do the following:
- Literature review on RL for pick-and-place maneuvers and multi-robot coordination
- Modeling of MARC in Isaac Sim
- Implementation of cuRobo for low-level collision-free robot arm path planning
- Training of an RL policy in Isaac Lab
- If time permits: deployment on real hardware
During your time at ARC, you will do the following:
- Literature review on RL for pick-and-place maneuvers and multi-robot coordination
- Modeling of MARC in Isaac Sim
- Implementation of cuRobo for low-level collision-free robot arm path planning
- Training of an RL policy in Isaac Lab
- If time permits: deployment on real hardware
Ideally, you already have the following skills or are eager to learn them:
- Proficiency in Python, ROS, and version control
- Autonomous working style
- Experience with Isaac Sim
- Experience with Reinforcement Learning for Robotics
Ideally, you already have the following skills or are eager to learn them:
- Proficiency in Python, ROS, and version control
- Autonomous working style
- Experience with Isaac Sim
- Experience with Reinforcement Learning for Robotics
Please send your CV, TOR and a short motivational statement to Jonas Stolle (jstolle@ethz.ch) and Emre Elbir (eelbir@ethz.ch)
Please send your CV, TOR and a short motivational statement to Jonas Stolle (jstolle@ethz.ch) and Emre Elbir (eelbir@ethz.ch)
Student Thesis - HS24 - Learning GPU Accelerated Multi-Robot Pick-And-Place.pdf
