|
|
||||||
 News
|
Research |
People |
WrightEagle |
Publications
|
Demo
|
Contact
|
|||||||
|
|
|
|
|
|
|
|
| |||||||||||||||||||||
| |||||||||||||||||||||
| During developing the WrightEagle team, we have done some demos. |
|
Download WMV MPG(High Quality) |
|
Download WMV MPG(High Quality) |
| These two demos show an example we did in a case study aiming at verifying Ke Jia's ability of planning for complex tasks. The complex tasks we chose are service queries of following form: "move a from position B to C and move b from D to E", given the initial state is shown in Fig. 1, where the robot is at location A and portable objects a and b are in position B and D, respectively. For a service robot that employs commands recognition technique, these two commands will be performed successively and separately, as shown in Figure 2. Generally this is not an optimal solution to the panning problem. |
| Ke Jia takes the user request as a single complex task and makes a plan interleaving the execution of the two component tasks this way: goto(B), pickup(a), goto(D), pickup(b), goto(C), putdown(a), goto(E), putdown(b), as shown in Figure 3. This plan is optimal with respect to the cost of Ke Jia's atomic actions, which comprises the distances between the locations/positions. More importantly, this plan was made autonomously and completely with general-purpose mechanisms. No matter whatever the locations/positions are chosen in the environment, Ke Jia is always able to generate an optimal plan for this complex task. This also means that Ke Jia can understand the interconnection among the atomic tasks contained in the complex task through her NLP mechanism. |
 Figure 1. The initial and the goal state of the complex task |
 Figure 2. Perform two component tasks separately |
 Figure 3. Perform the complex task optimally |
| In the experiments, Ke Jia was given the task in more natural forms, such as an English sentence like "give me the green bottle and put the red bottle on the table." The real-time performance is also acceptable. Generally, it took about 0.8 second for Ke Jia to accomplish the task panning. The computation of other modules is less time-consuming. |
|
Download WMV MPG(High Quality) |
| This video is the introduction of Ke Jia to her audience at RoboCup@home Competition at Graz, July, 2009. The name Ke Jia comes from an ancient Chinese tale, The Conch Fairy, which dates back to the 4th century. |
|
Download WMV MPG(High Quality)
|
| In this demo, Ke Jia was taught some causal knowledge through spoken dialog, with which she produced the same plan as one that was made with the build-in causal knowledge. In the experiments, we took a version of Ke Jia without any build-in knowledge about "balance", "fall", or any other equivalents. Ke Jia was told in the dialog that an object will fall if it is on the sticking-out end of a board and there is nothing on the other end of the board. With the knowledge and information, Ke Jia accomplished the task of catching the green bottle with the same plan: moving the red bottle first and them catching the green bottle. This indicates that Ke Jia's ability is substantially raised by knowledge acquisition through spoken dialog. It took no more than 0.5 second for Ke Jia to generate the plan with the taught knowledge. |
|
Download WMV MPG(High Quality) |
| It is well-known that the capability of reasoning with causation is very important and even crucial for many real-world applications. The causal reasoning problem we used in this case study is related to commonsense knowledge about "balance" and "fall". A testing instance is shown in Figure 1. There is a board putting on the edge of a table, with one end sticking-out. A red bottle is put on the sticking-out end of the board, and a green bottle is on the other end. If the green bottle was moved while the red one was on the stick-out end of the board, the red bottle would drop to the ground. The task is to pick up the green bottle under a default presupposition of avoiding anything falling. To accomplish the task, one must take some measure first against the unwanted side-effect of action "catch the green bottle". We input into Ke Jia some causal knowledge which is manually programmed and told Ke Jia that the red bottle is on the sticking-out end of the board and the green one is on the other end. Ke Jia produced a plan in which the red bottle is moved to a safety place first and then the green bottle is caught. The inference for generating the plan costs no more than 0.5 second in the experiments. The result demonstrates that Ke Jia can make use of build-in causal knowledge to accomplish causal reasoning easily. |
|
Recommended Web Browser for this page: Firefox |
| 2003-2009 © Multi-Agent Systems Lab, USTC | |