Sharing learned information among robots
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for sharing learned information among robots. In some implementations, a robot obtains sensor data indicating characteristics of an object. The robot determines a classification for the object and generates an embedding for the object using a machine learning model stored by the robot. The robot stores the generated embedding and data indicating the classification for the object. The robot sends the generated embedding and the data indicating the classification to a server system. The robot receives, from the server system, an embedding generated by a second robot and a corresponding classification. The robot stores the received embedding and the corresponding classification in the local cache of the robot. The robot may then use the information in the cache to identify objects.
1. A computer-implemented method comprising:
receiving, over a communication network by a server system from a first robot, a first embedding for a first object generated using a machine learning model stored by the first robot;
receiving, over the communication network by the server system from the first robot, data indicating a classification for the first object, wherein the first embedding and the data indicating the classification for the first object are stored in a local cache of the first robot; and
sending, over the communication network by the server system to the first robot, a second embedding and a corresponding classification for storage in the local cache of the first robot, wherein the second embedding is generated by a second robot and corresponds to a second object.
2. The computer-implemented method of claim 1 , further comprising providing, by the server system to the first robot, an updated machine learning model to replace the machine learning model stored by the first robot.
3. The computer-implemented method of claim 2 , wherein updated machine learning models are provided by the server system to the first robot periodically at a first interval larger than a second interval at which embeddings from one or more other robots are provided to the first robot.
4. The computer-implemented method of claim 2 , wherein the updated machine learning model is determined based on embeddings received by the server system from one or more other robots.
5. The computer-implemented method of claim 1 , further comprising:
receiving, by the server system, a plurality of embeddings from a plurality of robots; and
transmitting, by the server system to the first robot, aggregated embeddings comprising the plurality of embeddings received from the plurality of robots.
6. The computer-implemented method of claim 1 , further comprising storing, by the server system, a database including a table of classification labels corresponding to embeddings generated by a plurality of robots.
7. The computer-implemented method of claim 6 , wherein the database further comprises sensor data used to generate embeddings stored in the database.
8. The computer-implemented method of claim 6 , further comprising determining whether to share a particular embedding with one or more robots based on comparing a classification label received with the particular embedding with classification labels in the database.
9. The computer-implemented method of claim 1 , further comprising:
labelling, by the server system, additional embeddings with corresponding robot type; and
selectively sharing, by the server system, the additional embeddings with robots based on the corresponding robot type associated with each embedding.
10. The computer-implemented method of claim 1 , further comprising performing, by the server system, periodic synchronizations in which the server system requests each of a plurality of robots to upload any new embeddings since a last synchronization.
11. The computer-implemented method of claim 1 , further comprising determining, by the server system, which of a plurality of robots should receive one or more shared embeddings based on a role currently assigned to each of the plurality of robots.
12. The computer-implemented method of claim 1 , further comprising evaluating, by the server system, a quality of the data received from the first robot before distributing the data to one or more other robots.
13. The computer-implemented method of claim 1 , wherein the machine learning model is a neural network model, and wherein the first embedding is derived from output at an output layer of the neural network model.
14. The computer-implemented method of claim 1 , wherein the machine learning model is a neural network model, and wherein the first embedding is derived from the data indicating activations at a hidden layer of the neural network model.
15. A server system comprising:
one or more processors; and
one or more machine-readable media storing instructions that, when executed by the one or more processors, cause the server system to perform operations comprising:
receiving, over a communication network by the server system from a first robot, a first embedding for a first object generated using a machine learning model stored by the first robot;
receiving, over the communication network by the server system from the first robot, data indicating a classification for the first object, wherein the first embedding and the data indicating the classification for the first object are stored in a local cache of the first robot; and
sending, over the communication network by the server system to the first robot, a second embedding and a corresponding classification for storage in the local cache of the first robot, wherein the second embedding is generated by a second robot and corresponds to a second object.
16. The server system of claim 15 , the operations further comprising providing, by the server system to the first robot, an updated machine learning model to replace the machine learning model stored by the first robot.
17. The server system of claim 16 , wherein updated machine learning models are provided by the server system to the first robot periodically at a first interval larger than a second interval at which embeddings from one or more other robots are provided to the first robot.
18. The server system of claim 16 , wherein the updated machine learning model is determined based on embeddings received by the server system from one or more other robots.
19. The server system of claim 15 , the operations further comprising:
receiving, by the server system, a plurality of embeddings from a plurality of robots; and
transmitting, by the server system to the first robot, aggregated embeddings comprising the plurality of embeddings received from the plurality of robots.
20. One or more machine-readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
receiving, over a communication network by a server system from a first robot, a first embedding for a first object generated using a machine learning model stored by the first robot;
receiving, over the communication network by the server system from the first robot, data indicating a classification for the first object, wherein the first embedding and the data indicating the classification for the first object are stored in a local cache of the first robot; and
sending, over the communication network by the server system to the first robot, a second embedding and a corresponding classification for storage in the local cache of the first robot, wherein the second embedding is generated by a second robot and corresponds to a second object.