Update of local features model based on correction to robot action
Methods, apparatus, and computer-readable media for determining and utilizing corrections to robot actions. Some implementations are directed to updating a local features model of a robot in response to determining a human correction of an action performed by the robot. The local features model is used to determine, based on an embedding generated over a corresponding neural network model, one or more features that are most similar to the generated embedding. Updating the local features model in response to a human correction can include updating a feature embedding, of the local features model, that corresponds to the human correction. Adjustment(s) to the features model can immediately improve robot performance without necessitating retraining of the corresponding neural network model.
1. A method implemented by one or more processors of a robot, the method comprising:
storing, locally on one or more computer readable media of the robot, a past sensor data, feature instance, the past sensor data, feature instance comprising:
sensor data that captures an object, in an environment of the robot, and that is generated by at least one sensor of the robot, and
a feature of the object, wherein the feature is determined based on user interface input, and wherein the feature is stored as a feature of the object based on the user interface input being received in association with the sensor data that captures the object;
subsequent to storing the sensor data, feature instance:
receiving, via one or more network interfaces of the robot:
a new neural network model to replace an old neural network model locally stored on one or more computer readable media of the robot, and
a new features model to replace an old features model locally stored on one or more of the computer readable media of the robot;
in response to receiving the new neural network model and the new features model:
locally adapting the new features model based on the sensor data, feature instance; and
locally using the new neural network model and the new features model, as adapted, in controlling the robot and in lieu of the old neural network model and the old features model.
2. The method of claim 1 , wherein locally adapting the new features model based on the sensor data, feature instance comprises:
applying the sensor data as input to the new neural network model;
generating, over the new neural network model based on the applying, an embedding of the sensor data; and
determining, based on the embedding of the sensor data, a new feature embedding, of the feature, in the new features model.
3. The method of claim 2 , wherein determining, based on the embedding of the sensor data, the new feature embedding, of the feature, in the new features model, comprises:
identifying, in the new features model, a current feature embedding of the feature; and
determining the new feature embedding based on the embedding of the sensor data and the current feature embedding.
4. The method of claim 3 , wherein determining the new feature embedding based on the embedding of the sensor data and the current feature embedding comprises:
determining the new feature embedding based on an average of the embedding of the sensor data and the current feature embedding.
5. The method of claim 2 , further comprising:
prior to receiving the new neural network model and the new features model:
locally adapting the old features model based on the sensor data, feature instance.
6. The method of claim 5 , wherein locally adapting the old features model based on the sensor data, feature instance comprises:
applying the sensor data as input to the old neural network model;
generating, over the old neural network model based on the applying, an old embedding of the sensor data;
determining, based on the old embedding of the sensor data, an old feature embedding, of the feature, in the old features model.
7. The method of claim 1 , further comprising:
prior to receiving the new neural network model and the new features model:
locally adapting the old features model based on the sensor data, feature instance.
8. The method of claim 1 , wherein the sensor data is vision sensor data and the at least one sensor is a vision sensor.
9. A robot comprising:
one or more network interfaces;
one or more microphones;
a vision sensor;
memory storing instructions;
one or more processors executing the instructions, stored in the memory, to cause one or more of the processors to:
receive, via the one or more network interfaces of the robot:
a new neural network model, and
a new features model;
in response to receiving the new neural network model and the new features model:
locally adapt the new features model based on a locally stored vision sensor data, feature instance, wherein the vision sensor data, feature instance was locally stored prior to receiving the new neural network model and the new features model, and wherein the vision sensor data, feature instance comprises:
vision sensor data that captures an object, in an environment of the robot, and that is generated by the vision sensor, and
a feature of the object, wherein the feature is determined based on user interface input via the one or more microphones, and wherein the feature is stored as a feature of the object based on the user interface input being received in association with the vision a sensor data that captures the object; and
locally use the new neural network model and the new features model, as adapted, in controlling the robot.
10. The robot of claim 9 , wherein in locally adapting the new features model based on
the vision sensor data, feature instance, one or more of the processors are to:
apply the vision sensor data as input to the new neural network model;
generate, over the new neural network model based on the applying, an embedding of the vision sensor data; and
determine, based on the embedding of the vision sensor data, a new feature embedding, of the feature, in the new features model.
11. The robot of claim 10 , wherein in determining, based on the embedding of the sensor data, the new feature embedding, of the feature, in the new features model, one or more of the processors are to:
identify, in the new features model, a current feature embedding of the feature; and
determine the new feature embedding based on the embedding of the vision sensor data and the current feature embedding.
12. The robot of claim 11 , wherein in determining the new feature embedding based on the embedding of the vision sensor data and the current feature embedding, one or more of the processors are to:
determine the new feature embedding based on an average of the embedding of the vision sensor data and the current feature embedding.
13. The robot of claim 10 , wherein in executing the instructions one or more of the processors are to:
prior to receiving the new neural network model and the new features model:
locally adapt an old features model based on the vision sensor data, feature instance, wherein the new features model replaces the old features model.
14. The robot of claim 9 , wherein in executing the instructions one or more of the processors are to:
prior to receiving the new neural network model and the new features model:
locally adapt an old features model based on the vision sensor data, feature instance, wherein the new features model replaces the old features model.
15. The robot of claim 14 , wherein in locally adapting the old features model based on the vision sensor data, feature instance, one or more of the processors are to:
apply the vision sensor data as input to the old neural network model;
generate, over the old neural network model based on the applying, an old embedding of the vision sensor data;
determine, based on the old embedding of the vision sensor data, an old feature embedding, of the feature, in the old features model.
16. The robot of claim 14 , wherein the feature is a classification.
17. The robot of claim 14 , wherein the feature is a bounding area.