Skip to content
Vanessa Hassouna edited this page Aug 8, 2023 · 1 revision
  • Cram is responsible for 1. define which actions to execute to achieve the goal, 2. infer which parameters to use for each actionm, 3. monitor task execution and react to failures
  • Generlized plans: We call a robot action plan generlized if the same plan can accomplish a large amount of variations of one action category.
  • Scalability is achieved by underspecified symbolic action description.
  • Reactive planning approach, such that the plans perform the actions and monitor their execution concurrently.
  • Finding suitable action parameters -> a parameter is considered suitable if it results in successful execution of the action, first, in simulation and then also on the real robot.
  • Experience logs = episodic memories, they contain not only the low-level data streams, such as joint positions or camera images, but also the semantically annotated steps of the plan and their relations.
  • To perceive an object in the environment, the plan executive send a vague object description to the perception system.
  • Motion Planning send a goal-poses, this can contain poses for different end effectors, e.g., for a dual-arm motion.
  • The collision-mode can be any of: avoid-all, allow-all, allow-arm, allow-hand, allow-fingers or allow-fingers-and-objects, which specifies if collisions are allowed between the robot body parts or the grasped object.
  • Goals are defined as a task function f(o) with limits its first derivative. The vector of observable variables, o, contains the robot joint positions, but can also contain goal parameters, poses of environment objects, contact information, etc.
Clone this wiki locally