Loading...
Please wait, while we are loading the content...
Bayesian Contact State Segmentation for Programming by Human Demonstration in Compliant Motion Tasks
| Content Provider | Semantic Scholar |
|---|---|
| Author | Meeussen, Wim Rutgeerts, Johan Gadeyne, Klaas Bruyninckx, Herman Schutter, Joris De |
| Copyright Year | 2006 |
| Abstract | This paper presents a contribution to programming by human demonstration [3], [14], in the context of compliant motion [4] tasks in which an object held by a manipulator moves in contact with an environmental object, as shown in Fig. 1. Major challenges in the automatic translation from a human compliant motion demonstration into an executable compliant motion robot program are: (i) to recognize the contact formation (CF) to which the human demonstration is currently subjected, (ii) to estimate the geometric parameters of that contact formation, and (iii) to detect when exactly the human demonstration execution changes between two CFs. Initial research on the identification of CFs focused mainly on two different approaches: (i) ad hoc identification strategies that exploit geometric knowledge of the contacting objects, but that have a very poor stochastic foundation (e.g. [1], [7]) and (ii) Hidden Markov Model based (hence stochastic) solutions to assembly problems that can recognize CF transitions very fast but only with limited allowed uncertainty (e.g. [6], [12]). Slaets et al. [13] presented some results in the field of compliant motion that go a bit beyond the scope of this paper: they build a geometric model of an unknown environment from a given number of primitives, while this paper starts from a known parameterization of a geometric model. However, their approach is based on the Non-Minimal State Kalman Filter [9], [11], and is only valid if the estimation has converged to a unimodal Gaussian before a contact transition. They only allow new contact constraints to be added gradually, and no contact constraints to be removed. Recently Gadeyne et al. [8] developed a particle filter [5] approach to estimate, simultaneously, geometric parameters of a known geometric model and to recognize contact transitions. Their approach is able to estimate (continuous) geometric parameters with a large uncertainty, and simultaneously recognize (discrete) contact transitions in an experiment consisting of six possible and initially known CFs. This paper generalizes and scales the approach of Gadeyne et al. to allow all possible contacts between two polyhedral objects. To cope with this increased complexity, an improved prediction step is used, based on the topological information contained in a contact state graph, [15], [16], and the relative pose of the contacting objects. This paper also presents efficient algorithms for the pose and consistency measurement equations [2], [10], that reduce the numerical cost and allow the estimators to discriminate in realtime between 245 different CFs, in an uncertain environment. |
| Starting Page | 3 |
| Ending Page | 12 |
| Page Count | 10 |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://people.mech.kuleuven.be/~wmeeusse/Papers/iser2006_abstract.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
