Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Learning grasping reflexes with imperfect depth sensors
Manipulators potentially extend the functionality of legged robots. This project aims to explore methods to improve the robustness of the grasping policies.
Keywords: Grasping; reinforcement learning.
Deploying the manipulators on the legged platforms demands robustness to dynamics disturbances and sensor limitations: due to the stepping gaits of the legs, tracking the end-effector with high accuracy is difficult; the robot may not fully observe the target object when the gripper moves close due to the limited field-of-view. Previous work in locomotion has demonstrated that reinforcement learning can produce proprioceptive reflexes [1] and internal belief of the environment states [2]. This project aims to explore deploying such methods to grasp known objects under dynamics noise and limited perception.
The result will be validated on the Dynaarm and potentially on the whole anymal+arm system in the lab.
[1] J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,” Science robotics, vol. 5, no. 47, 2020.
[2] T. Miki, J. Lee, J. Hwanbo, L. Wellhausen, V. Koltun, and M. Hutter, “Wild anymal: Learning robust perceptive locomotion,” Under review, 2021.
Deploying the manipulators on the legged platforms demands robustness to dynamics disturbances and sensor limitations: due to the stepping gaits of the legs, tracking the end-effector with high accuracy is difficult; the robot may not fully observe the target object when the gripper moves close due to the limited field-of-view. Previous work in locomotion has demonstrated that reinforcement learning can produce proprioceptive reflexes [1] and internal belief of the environment states [2]. This project aims to explore deploying such methods to grasp known objects under dynamics noise and limited perception.
The result will be validated on the Dynaarm and potentially on the whole anymal+arm system in the lab.
[1] J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,” Science robotics, vol. 5, no. 47, 2020.
[2] T. Miki, J. Lee, J. Hwanbo, L. Wellhausen, V. Koltun, and M. Hutter, “Wild anymal: Learning robust perceptive locomotion,” Under review, 2021.
- Literature review and familiarizing with the software infrastructure
- Implementing the training environment
- Training the grasping policy
- Evaluation of the policy and the training methods
- Validating the trained policy on the hardware
- Literature review and familiarizing with the software infrastructure - Implementing the training environment - Training the grasping policy - Evaluation of the policy and the training methods - Validating the trained policy on the hardware
- High motivation and interest in the topic
- Good Python programming skills
- Knowledge of (deep) reinforcement learning
- Experience with machine learning tools (Pytorch or Tensorflow)
- Have taken machine learning courses
- Having taken the Robot Dynamics course would be a plus
- High motivation and interest in the topic - Good Python programming skills - Knowledge of (deep) reinforcement learning - Experience with machine learning tools (Pytorch or Tensorflow) - Have taken machine learning courses - Having taken the Robot Dynamics course would be a plus
Please send your CV and transcript to Yuntao Ma (mayun@ethz.ch)
Please send your CV and transcript to Yuntao Ma (mayun@ethz.ch)