Skip to content

Latest commit

 

History

History
183 lines (126 loc) · 5.63 KB

README.md

File metadata and controls

183 lines (126 loc) · 5.63 KB
Raw

AffAction

Affordances and actions for robot behavior generation.

Concepts

Conceptually, this project connects a semantic text-level action language with the low-level motor control of a robot (or its simulation).

We represent a scene as a composition of several objects, each of them comprising several affordances. For instance, a table can have the "Supportable" affordance to indicate that something can be put on its support surface. Similarly, objects can have the Affodance to be "Graspable", where this can be detailed down to "Powergraspable" for cylindrical objects etc. Affordance come typically with a coordinate frame that describes the geometry of its interaction. For instance, the "Powergraspable" affordance is described by a frame that has the z-axis pointing into the invariant rotation direction.

The scene also contains a set of Manipulators that are composed of several "Capabilities". A "Capability" can for instance be a "Powergrasp", a "Palmgrasp", a "FingerpushCapability" and more. At runtime, an action is checking all possible combinations of Capabilities and Affordances, and computes the feasible pairs.

Here is an example: The action "get apple" leads to:

  • matching all manipulator's (e.g. left and right hand) capabilities (e.g. Powergraps, Palmgrasp) against all object affordances (e.g. how it is graspable). The resulting combinations are sorted according to a heuristic (usually distance in wrench space). In this computation step, the robot's particular embodiment (reachability, joint ranges ...) is not considered.
  • The action is simulated forward with a prediction class, and is checked against violations of the robot's limits and against collisiins. The first feasible solution is sent for execution. In this computation step, the robot's embodiment is considered.

Actions

The "get" action

get(<object-to-pick-up> <which-hand (optional)> <"from" parent-object (optional>)

Grasps an object and lifts it a little bit up. If no manipulator (which-hand) is given, the action will determine the best possible manipulator. Internally, the action matches the manipulator's capabilities with the object's affordances. These combinations are supprted:

Affordance Capability
PowerGraspable PowergraspCapability
PincerGraspable PincergraspCapability
BallGraspable PincergraspCapability
PalmGraspable PalmgraspCapability

The set of matches is evaluated, and the best feasible one is determned. If the "from" attribute is used, the object-to-pick-up is determined as being a child of the object-to-pick-up. This allows to give a spatial constraint in case there are several objects with the same name. For instance:

get lemon from cutboard   // Gets a lemon that is a child of the cutboard. The hand is auto-selected.
get apple_1 hand_right    // Gets the apple_1 with the right hand.

The "put" action

put(, <location-hint (optional)> <which-hand (optional)>)

The "double_get" action

double_get (<object1, object2>)

The "double_put" action

double_put(, , (target1), (target2))

The "poke" action

poke()

The "gaze" action

gaze()

The "open_door" action

open_door()

The "close_door" action

close_door ()

The "pour" action

pour(,)

The "pose" action

pose(<name-of-a-model-state-pose>)

This action loads the passed model state from the graph's configuration file and moves all joints into this pose. If a join is not part of the model state, it will not be moved.

The "push" action (Not yet working)

push()

The "screw" action (Not yet working)

screw()

Feedback messages

Syntax of Error messages:

ERROR: cannot  REASON: because  SUGGESTION:  DEVELOPER:

or

FATAL_ERROR for non-recoverable errors (e.g. Emergency Stop ...)

Unrecoverable Errors that Need a Replan:

Semantic Errors:

  • Wrong command syntax
  • Usage of object names not existing in the environment

Logical Errors:

  • Try to get an object when both hands are full
  • Try to pour an object into an object placed in a closed container
  • Try to pour an object into an object already full
  • Opening or closing non openable objects
  • Switching on/off non powerable objects

Physical Errors:

  • Command impossible to perform by the robot because of joint-limit errors
  • Command impossible to perform because object x is an obstacle
  • Command impossible to perform because target object is out of reach

Software design

The project is composed of three parts:

  • Actions: Algorithmic libraries that implement the ActionScene, Affordances, Capabilities and actions.
  • Component system: Components that interact with the subscribe - publish mechanism
  • Examples: Example classes that work with the ExampleRunner applications.

How to build

  • Please make sure that you have set the SIT and MAKEFILE_PLATFORM environment variables
  • Please make sure the WM5 library has been compiled into the Rcs dependency. If not, a runtime warning will be issued.
  git clone https://github.com/HRI-EU/affaction.git
  cd affaction
  mkdir build
  cd build
  cmake ..
  make
  • This should build several executables into the bin directory

How to start the websocket action server

  • bin/TestLLMSim -port 35000 (that's the default)
  • with command line options printed to console: bin/TestLLMSim -h

Python websocket client

  • cd SmileActions/python
  • python smile_websocket.py "put cola_bottle_1"