Archive for October, 2023
Robot Cooking – Transferring observations into a planning language
Transferring observations into a planning language: An automated approach in the field of cooking
In the Robot Cooking project, an automated method is developed to analyze and identify motion data and convert it into a machine-readable planning language. This is done using a cooking scenario as an example in which the motion data is captured by recording the hand pose of the cook.
The recording is done using a motion capture system consisting of seven cameras and a glove with three markers on the back of the chef’s hand. The position of the markers is determined by triangulation. This provides enough information to derive the hand pose. The recording is done at 120 frames per second. Before the cooking process, all objects in the workspace are identified and their initial positions determined. Motion data is continuously recorded and converted into poses with time stamps. Additional information such as velocity, acceleration and angle in relation to the tabletop are derived from the raw data.
A initial structure of the dataset is created by finding the side actions using classification. Here, pick, move and place are identified as recurrent side actions. A separate training dataset is used to train a classifier that recognizes these actions. This enables an easier analysis of the remaining actions.
Clustering is applied to identify unknown actions. A dynamic approach allows analysis despite high variability in execution. A unique fingerprint for each action is found, based on the orientation of the back of the hand and its speed on the table plane, to assign each frame to a cluster and finally to an action.
The knowledge gained from classification and clustering is translated into a machine-readable Planning Domain Definition Language (PDDL). A schedule is created, with known actions directly assigned. Start and end positions are specified, and virtual object tracking is used to represent the progression of objects during cooking. For unknown actions, preconditions and effects are handled dynamically. The results are translated into a machine-readable PDDL. This formal representation enables efficient automatic scheduling and execution of the previously demonstrated cooking task.
Additional information is available in the video linked above, the poster and the paper.
Contact:
Markus Schmitz
