Symbolic state estimation for contact-rich manipulation tasks
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for controlling a robot using symbolic states. One of the methods includes receiving a definition of a task having multiple task states, wherein each task state is associated with a different respective control policy; executing the task using an initial control policy associated with an initial task state; during execution of the task, continually generating, from sensor data, a prediction of the task state of the task; and upon determining that a transition to a different task state has occurred, transitioning the robot to a different control policy associated with the different task state.
1. A method for controlling a robot comprising:
receiving a definition of a task, wherein:
the task has multiple task states,
each task state is associated with a different respective control policy, and
each task state is defined by a composition of one or more predicate values;
executing the task using an initial control policy associated with an initial task state for the task;
during execution of the task:
continually generating, from sensor data, one or more predicted predicate values for each of one or more predicates defining the multiple task states, and
generating, from the one or more predicted predicate values and one or more task state transition prior belief states, a resulting prediction of the task state that the task is currently in; and
when a transition from the initial task state to a different task state has occurred, controlling the robot using a different control policy associated with the different task state of the task.
2. The method of claim 1 , wherein each predicate classifier of a plurality of predicate classifiers uses sensor data from only a single sensor.
3. The method of claim 1 , wherein a first predicate classifier uses only image data.
4. The method of claim 3 , wherein a second predicate classifier uses only robot position, velocity, and force data.
5. The method of claim 1 , wherein generating the prediction of the task state comprises using the one or more predicted predicate values as input to a model trained to generate a predicted task state from predicate compositions.
6. The method of claim 1 , wherein the task is an insertion task, and wherein the task states comprise an above task state representing that a connector is above a socket, an on-surface task state representing that the connector has contacted the surface of an object, an aligned task state representing that the connector is aligned with the socket, or an in task state that represents that the connector has been inserted.
7. The method of claim 1 , wherein the plurality of task states include one or more task states representing error conditions associated with the robot.
8. A system comprising:
one or more computers; and
one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
receiving a definition of a task, wherein:
the task has multiple task states,
each task state is associated with a different respective control policy, and
each task state is defined by a composition of one or more predicate values;
executing the task using an initial control policy associated with an initial task state for the task;
during execution of the task:
continually generating, from sensor data, one or more predicted predicate values for each of one or more predicates defining the multiple task states, and
generating, from the one or more predicted predicate values and one or more task state transition prior belief states, a resulting prediction of the task state that the task is currently in; and
when a transition from the initial task state to a different task state has occurred, controlling a robot using a different control policy associated with the different task state of the task.
9. The system of claim 8 , wherein each predicate classifier of a plurality of predicate classifiers uses sensor data from only a single sensor.
10. The system of claim 8 , wherein a first predicate classifier uses only image data.
11. The system of claim 10 , wherein a second predicate classifier uses only robot position, velocity, and force data.
12. The system of claim 8 , wherein generating the prediction of the task state comprises using the one or more predicted predicate values as input to a model trained to generate a predicted task state from predicate compositions.
13. The system of claim 8 , wherein the task is an insertion task, and wherein the task states comprise an above task state representing that a connector is above a socket, an on-surface task state representing that the connector has contacted the surface of an object, an aligned task state representing that the connector is aligned with the socket, or an in task state that represents that the connector has been inserted.
14. The system of claim 8 , wherein the plurality of task states include one or more task states representing error conditions associated with the robot.
15. One or more non-transitory computer storage media encoded with computer program instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
receiving a definition of a task, wherein:
the task has multiple task states,
each task state is associated with a different respective control policy, and
each task state is defined by a composition of one or more predicate values;
executing the task using an initial control policy associated with an initial task state for the task;
during execution of the task:
continually generating, from sensor data, one or more predicted predicate values for each of one or more predicates defining the multiple task states, and
generating, from the one or more predicted predicate values and one or more task state transition prior belief states, a resulting prediction of the task state that the task is currently in; and
when a transition from the initial task state to a different task state has occurred, controlling a robot using a different control policy associated with the different task state of the task.
16. The one or more computer storage media of claim 15 , wherein each predicate classifier of a plurality of predicate classifiers uses sensor data from only a single sensor.
17. The one or more computer storage media of claim 15 , wherein a first predicate classifier uses only image data.
18. The one or more computer storage media of claim 17 , wherein a second predicate classifier uses only robot position, velocity, and force data.
19. The one or more computer storage media of claim 15 , wherein generating the prediction of the task state comprises using the one or more predicted predicate values as input to a model trained to generate a predicted task state from predicate compositions.
20. The one or more computer storage media of claim 15 , wherein the task is an insertion task, and wherein the task states comprise an above task state representing that a connector is above a socket, an on-surface task state representing that the connector has contacted the surface of an object, an aligned task state representing that the connector is aligned with the socket, or an in task state that represents that the connector has been inserted.