[In this semester, I want to read many papers on different aspects of artificial intelligence concentrating on machine learning (in general), reinforcement learning (in particular), evolutionary mechanisms, intelligent robotics, and â¦ . In this way, I try to write my notes and comments on the papers that I read. I will submit my comments to the papersâ authors and see if I can get a good feedback.
I will find the optimal way of doing so gradually. Different parameters affect the behavior ranging from my time to the benefit of this work for myself. Any suggestion or comments is highly appreciated.]
Martin Jagersand, “Image Based Visual Simulation and Tele-Assisted Robot Control,” IROS 97.
Martin Jagersand suggested me to read a few of his papers. This is the first of that suggested series. In this paper, Martin develops a novel visual servo mechanism for a robotic manipulator. As I am not aware of the literature of visual servoing, I cannot compare his method with others and find a web of related work. Therefore, I try to assess it based on what it says and my own prior knowledge. In future, if I become more knowledgeable of machine vision, robotic manipulators, and similar subjects, I will judge it better.
The main and the most important idea of the paper, in my opinion, is doing the task in visual space and estimating visual-motor model of the system by estimating the Jacobian of the âjoint variablesâ to âfeatures that can be seenâ. It is interesting/important as doing so allows us not to estimate two models of world space->visual space and joint space->world space. Not only it is computationally advantageous, but also it might be more robust. There are some other important things in this paper such as using that model to predict the manipulatorâs movement in the visual world and the idea of virtual tool and using it to solve a complete task. In addition, there is emphasis on tele-assistance and the way human can guide the robot to do tasks.
Let me go to details and point out some important aspects of the paper:
The paper starts with this question: âWhat makes the human able to effortlessly move and perform manipulation in natural and non-engineering environments where robots fail?â and it answers ââ¦ while robot manipulation tends to rely maximally on pre-programmed models and minimally on sensory information, the human need rely very little on a priori models, and typically relies maximally on sensory informationâ. Do I agree with this? Somehow yes, somehow no! I believe that robots tend to rely on pre-programmed models and this over-exaggerated insistence prevents them to be robust, adaptive, and creative in novel situations (the root of this problem stems from classical control theory which swerves cybernetics from what it intended to be). Classical approach to intelligence tries to model the world as exact as possible. It is very good to perform precisely predefined tasks, but it fails to work in noisy, uncertain, or new environments. Now, situated embodied robotics appears and solves this problem in some extent. This research exactly follows this way. It uses real robot that, by benefiting the interaction between itself and the environment, estimates the systemâs model. However, this situated and embodied research track may not be that sufficient to explain the human competence. I believe that there is a need in different layers of abstraction to confront novel and difficult situations. Apparently, this belief stems from my interests in hierarchical structures and my probable research on abstraction.
+The idea of defining a model agreement (used in (8)) is interested for me. I need to draw some curves with positive and negative second derivative of J_true to understand its meaning. It might be possible to estimate H_inf norm of J_error (=J_true â J_estimated) using deltaY_estimated and delta_X and then, use it to make a robust controller. However, I am not sure whether it is significantly better.
+The paper uses an adaptive mechanism to change J. However, J is configuration dependent and its changes are due to both environment uncertainty and system changes (e.g. when the robot picks a box) and its configuration in space. Adapting a model that changes intrinsically is possible (as the paper reports), but it is not the best solution in my opinion. If I were him, I use a multi-model controller, i.e. something like adaptive gain scheduler. The system starts with a random (or some predefined) visual-motor estimation. When it moves along, it updates the model. However, when it moves sufficiently far, it stores the model in a model toolbox and makes a copy of the previous model and updates it afterward. This process continues for ever with some maintenance mechanism to prevent the toolbox to become too big. Whenever the robot comes close to the previously met state in the visual-space, it uses one of those acquired-before models. If the dynamics of the system were not changed in that duration, this new model acts very precisely.
+I wonder if it is possible to add learning to this problem and use tele-assistance as external critic to assess the performance. Primitive movements such as âalignmentâ can be considered as options and therefore, their sequencing can be learnt.
+It is interesting for me to know whether the estimation under control action generates persistent exciting (PE) vectors for Jacobian estimator (6) (Equation (6) seems to be a kind of projection algorithm). I think that the natural noise in the robotic task make the signals PE.