16.08.2021 Meeting Notes

Vedant Dave notes_vedant
image_pdfimage_print

Meeting Details

Date : 13th August 2021

Time : 14:30 – 15:30

Location : Chair of CPS, Montanuniverität Leoben

Participants: Univ.-Prof. Dr. Elmar Rueckert, Vedant Dave

Agenda

PhD Thesis direction initiation

Topic 1: Improvement in Riemannian Approach

  1. Closed form solution for the Cost function instead of Gradient Descent.
  2. Decreasing computational time algorithmically or with parallel computing.
  3. ProMPs on Mixture Manifolds: ℝ3 × 𝕊3.

Topic 2: Topics of Possible Research Directions

  1. Conditional Neural Movement Primitives
    • Implement the initial approach.
    • Develop this approach for Orientation and full ProMPs.
    • The part of Control is missing from the paper, which can be worked upon.
    •  
  2. ProMPs with Reinforcement Learning
    • Use Reinforcement Learning to learn the model outside the demonstrated trajectory distribution.
    • Reinforcement Learning for generalizing to new configurations.
        •  
  3.  ProMPs wit Transfer Learning
    •  Search the research papers focusing on both, Movement Primitives and Transfer learning. The advantage lies in learning from minimal training data.
    • Look up papers for combining the Model-agnostic Meta learning with Movement Primitives.
        •  
  4.   Planning as Inference
    • Read up for Research paper concerning the topic.
    • Online trajectory modulation.
        •  
  5. Human-Machine Interaction and Control
      •  
  6. ProMPs with latent manifolds
      •  
  7. Constrained Probabilistic Movement Primitives
    • Understand the approach of the recent paper and check for any other possible improvement and possible research direction.
  •  

Next Steps

  1. Search for relevant research papers in those fields.
  2. Finalise the first project to work on.
  3.  Meet and discuss the topic with Nicolas. Also make the topic adaptable to his Master Thesis.

Literature

ProMPs with Latent Manifolds

  1. Rückert EA, Neumann G. Stochastic optimal control methods for investigating the power of morphological computation. Artif Life. 2013 Winter;19(1):115-31. doi: 10.1162/ARTL_a_00085. Epub 2012 Nov 27. PMID: 23186345. 

Conditional Neural Movement Primitives

  1. Seker, M., Imre, M., Piater, J., & Ugur, E. (2019). Conditional Neural Movement Primitives. Robotics: Science and Systems
  2. Garnelo, Marta & Rosenbaum, Dan & Maddison, Christopher & Ramalho, Tiago & Saxton, David & Shanahan, Murray & Teh, Yee & Rezende, Danilo & Eslami, S.. (2018). Conditional Neural Processes. 

ProMPs wit Transfer Learning

  1. Stark, S., Jan Peters and Elmar Rueckert. “Experience Reuse with Probabilistic Movement Primitives.” 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2019): 1210-1217.
  2. Finn, Chelsea, P. Abbeel and Sergey Levine. “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks.” ICML (2017).  

ProMPs with constraints

  1. Frank, F., Paraschos, A., Smagt, P.V., & Cseke, B. (2021). Constrained Probabilistic Movement Primitives for Robot Trajectory Adaptation. ArXiv, abs/2101.12561.

Next Meeting

Next meeting was scheduled on Tuesday, 17th August, 2021. 11:00 am – 12:00 pm.