Stabilizing novel objects by learning to predict tactile slip
Veiga, F., van Hoof, H., Peters, J., & Hermans, T. (2015). Stabilizing Novel Objects by Learning to Predict Tactile Slip. Proceedings Of the IEEE/RSJ Conference on Intelligent Robots and Systems. Retrieved from http://www.ausy.tu-darmstadt.de/uploads/Site/EditPublication/IROS2015veiga.pdf
Summary
Veiga et al. compare a variety of supervised machine learning methods for learning to detect and predict tactile slip. They compare both SVMs and random forest classifiers and use features that come from a tactile sensor based on the human finger. More specifically, they have three different ways of constructing features:
- Single element features – based only on the current timestep, $\phi(\mathbf{x}_{1:t})=\mathbf{x}_t$
- Delta features – based on the current timestep, as well as the change from the last timestep, $\phi(\mathbf{x}_{1:t})=[\mathbf{x}_t,\Delta\mathbf{x}_t]$
- Time window features – based on a window of previous time steps, $\phi(\mathbf{x}_{1:t})=\mathbf{x}_{t-\tau:t}$, where $\tau$ is the size of the time window
The random forest classifiers tend to do the best, particularly with the delta features.
Takeaways
In this paper, Veiga et al. use purely discriminative methods to detect and predict slip based on features extracted from a tactile sensor. This means that their method works independently of object-specific features and thus does not take into account anything about the actual dynamics of the object. In this case, this approach actually makes a lot of sense to me—it feels like it would be overkill to need to have a full dynamics model of how the object behaves just to determine if it is slipping. That said, I do wonder if having some approximate knowledge of the object dynamics (e.g. curvature, overall mass, distribution of mass, etc.) would help, particularly in having the controller adjust once it has detected that the object is slipping.
I also wonder to what extent these learned classifiers will generalize to greater changes in mass/shape/friction—e.g. will the same classifier work for a very heavy object as opposed to a very light object? Or a very smooth object vs. a rough object? And, will it generalize to other types of scenarios where slip might occur (e.g. holding an object with two hands, as opposed to up against a wall; and also, the location where the object is being held).
This paper also made me think a bit about what constitutes simulation. In this paper, Veiga et al. are able to predict several timesteps into the future whether slip will occur. Even though they don’t have a dynamics model of how slip changes over time, for example, does this type of prediction still count as “simulation”? I’m tempted to say no, not unless they can predict at each timestep how much slip is going to occur—e.g. in the next timestep, there will be a small amount of slip, in the following timestep, it will increase by so much, etc. I don’t think this type of prediction would be particularly hard to do, though, if they could perhaps attach sensors to their objects in order to more precisely quantify what slip is (in this paper, they labeled parts of the video as “slip” or “not slip” just by human visual judgments).