Theme

Last Updated: 2000.6.7

Theme

Learning from Observation

Researches (Abstract)

APO(Assembly Plan from Observation)

Analysis of assembly tasks, which are composed of polyhedral objects, based on contact-state transitions.
It acquires the trajectory of the manipulated objects by 3D tracking.
Then it analyzes the changes of contact-state transitions from the trajectory and represents the task as a finite contact-state transitions.
By assigning a sub-skill to each contact-state transition which produces the expected transition, the robot can peformes the same task.

AP(Attention Point) Analysis

Proposition of Attention Point:"Points (time and position) in an observation of a task which demands detailed analysis later."
Sequential refinement of a task model by locally analyzing observations based on attention point analysis.
- Spatial integration of multiple sensors
  1st step: By using some low-cost sensors(EX data-gloves), it constructs rough task model and extracts attention points which should be analyzed in detail later. The input data of un-processed sensors are recorded on a storage device.
  2nd and later steps: Remaining sensors (EX stereo images) which corresponds to the attention points are fetched and analyzed in detail, and the locally enhanced task model is generated sequentially.
- Temporal integration of multiple observations
  1st observation: Sensors are configured (EX x1 zoom stereo, a rough model) to observe the entire human demonstration, then the system constructs a rough task model and extracts attention points.
  2nd and later observations: The same sensors are configure in different way (EX x2 zoom stereo, a fine model) to observe the partial hman demonstration around the attention points in detail and the system generates a locally enhanced task model.

Real-time 3D Tracking

With a real-time stereo vision system(30fps, 280x200 pixels), the system localizes objects in the 3D space fast by 3DTM(3D Template Matching) technique.