Learning probability distributions over partially-ordered human everyday activities. Tenorth, De la Torre, Beetz. ICRA 2013

  1. Attempt “… to learn the partially ordered structure inherent in human everyday activities from observations by exploiting variability in the data.”
  2. Learns full joint probability over actions that make up a task, their partial ordering, and their parameters
  3. Can be used for classification, but also figuring out what actions are relevant to a task, what objects are used, if it was done correctly, or what is typical for an individual
  4. Use synthetic data as well as TUM Kitchen and CMU MMAC
  5. Use Bayesian Logic Nets (another paper has author overlap and uses the same approach)
  6. Common alternate approaches are HMMs, conditional random fields (CRFs), or suffix trees
    1. But these are most effective when the ordering of the subtasks are pretty fixed
    2. Also Markov assumption of HMM doesn’t really hold in the way data is often represented and may require all history information
  7. Also some other approaches for representing partial ordering
    1. <Whatever this means> “All these approaches focus only on the ordering among atomic action entities, while our system learns a distribution over the order as well as the action parameters.”
  8. Literature on partially ordered plans require lots of prior information, and have been applied to synthetic problems
  9. Working off the TUM dataset, they needed to resort to bagging and data noisification techniques in order to get enough samples
  10. Needs annotation / data labeling
  11. Learns, for example, that after making brownies (the CMU MMAC dataset) some people like to put the frying pan in the sink, and others on the stove
  12. Performance of this approach is much better than that of CRFs, and is more robust to variability in how people undertake tasks

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: