Robotic device control optimization using spring lattice deformation model
Method and apparatus for training a machine learning model for controlling a robotic picking device having a suction device end effector. A robotic control operation and a candidate contact point for holding a first item using a suction device of a robotic picking arm are determined, by processing information describing the first item as an input to a machine learning model. A seal quality metric is estimated for the candidate contact point, based on a predicted deformed n-dimensional shape of the suction device and a n-dimensional shape associated with the first item. One or more weights within the machine learning model are refined based on the estimated seal quality metric.
1. A computer-implemented method performed via execution of one or more computer processors, the computer-implemented method, comprising:
determining, by at least one of the one or more computer processors processing a first set of inputs using a machine learning model for determining behavior of a robotic arm, a first behavior for retrieving an item from within a storage container by holding the item at a contact point using a suction device of the robotic arm;
estimating, by at least one of the one or more computer processors, a likelihood of successfully retrieving the item using the determined first behavior, by determining an expected seal quality metric for the contact point, by processing an n-dimensional surface model of the item and physical properties of the suction device of the robotic arm, wherein determining the expected seal quality metric comprises:
modelling deformation and contact between the suction device and the n-dimensional surface model at the contact point on the item, using at least one n-dimensional spring-lattice to model deformation and contact between the suction device and the n-dimensional surface model at the contact point, such that potential energy across the n-dimensional spring-lattice is minimized; and
calculating the expected seal quality metric, using the n-dimensional spring-lattice and a fluids model relating fluid force and leak pathways; and
training, by at least one of the one or more computer processors, the machine learning model to optimize the behavior of the robotic arm, based on the expected seal quality metric.
2. The computer-implemented method of claim 1 , wherein training the machine learning model to optimize the behavior of the robotic arm comprises upon determining that the expected seal quality metric does not exceed a predefined threshold level of seal quality, refining one or more weights within the machine learning model to reduce a correlation between the first set of inputs and the first behavior.
3. The computer-implemented method of claim 1 , wherein training the machine learning model to optimize the behavior of the robotic arm comprises upon determining that the expected seal quality metric exceeds a predefined threshold level of seal quality, refining one or more weights within the machine learning model to increase a correlation between the first set of inputs and the first behavior.
4. The computer-implemented method of claim 1 , further comprising dynamically generating, by at least one of the one or more computer processors, the first set of inputs using a synthetic input generation engine.
5. The computer-implemented method of claim 4 , wherein the first set of inputs includes at least one of (i) robotic sensor data, (ii) item metadata, (iii) surface geometry information and (iv) suction device metadata.
6. The computer-implemented method of claim 4 , further comprising training, by at least one of the one or more computer processors, the synthetic input generation engine based on a set of historical input data observed within a fulfillment center environment.
7. A system, comprising:
one or more computer processors; and
computer program code that, when executed by operation of the one or more computer processors, performs an operation comprising:
determining a robotic control operation and a candidate contact point for holding a first item using a suction device of a robotic picking arm, by processing information describing the first item as an input to a machine learning model;
estimating a seal quality metric for the candidate contact point, based on a predicted deformed n-dimensional shape of the suction device and a n-dimensional shape associated with the first item; and
refining one or more weights within the machine learning model based on the estimated seal quality metric.
8. The system of claim 7 , wherein refining the one or more weights within the machine learning model based on the estimated seal quality metric comprises:
determining whether the estimated seal quality metric satisfies a first predefined criteria by determining whether the estimated seal quality metric exceeds a predefined threshold level of seal quality; and
refining the one or more weights within the machine learning model, based on whether the estimated seal quality metric satisfies the first predefined criteria.
9. The system of claim 7 , wherein estimating the seal quality metric for the candidate contact point, based on the predicted deformed n-dimensional shape of the suction device and the n-dimensional shape associated with the first item comprises
generating the predicted deformed n-dimensional shape of the suction device by modelling deformation and contact between the suction device and the n-dimensional shape associated with the first item at the candidate contact point, using at least a first n-dimensional spring-lattice to model deformation and contact between the suction device and the n-dimensional shape associated with the first item at the fir candidate contact point, such that potential energy across the first n-dimensional spring-lattice is minimized.
10. The system of claim 9 , wherein estimating the seal quality metric for the candidate contact point, based on the predicted deformed n-dimensional shape of the suction device and the n-dimensional shape associated with the first item further comprises
calculating the seal quality metric, using the first n-dimensional spring-lattice and a fluids model relating fluid force and leak pathways.
11. The system of claim 10 , wherein estimating the seal quality metric for the candidate contact point, based on the predicted deformed n-dimensional shape of the suction device and the n-dimensional shape associated with the first item further comprises:
retrieving an item profile corresponding to the first item and specifying data describing physical attributes of the first item; and
generating a second predicted deformed n-dimensional shape for the first item, using at least a second n-dimensional spring-lattice to model deformation and contact between the suction device and the predicted deformed n-dimensional shape for the first item at the candidate contact point, such that potential energy across the first n-dimensional spring-lattice is minimized,
wherein calculating the seal quality metric is further based on the second predicted deformed n-dimensional shape for the first item.
12. The system of claim 7 , the operation further comprising:
training a synthetic input generation engine based on a set of historical input data observed within a fulfillment center environment; and
dynamically generating at least one input to the machine learning model using the synthetic input generation engine.
13. The system of claim 7 , wherein the information describing the first item comprises a n-dimensional model of the first item and a determined pose of the first item.
14. The system of claim 13 , wherein the information describing the first item further comprises metadata describing physical attributes of the first item.
15. A non-transitory computer-readable medium containing computer program code that, when executed by operation of one or more computer processors, performs an operation comprising:
dynamically generating a plurality of synthetic machine learning model input data using a synthetic input generation engine;
determining, by processing the plurality of synthetic machine learning model input data as inputs to a machine learning model for determining control behavior for a robotic arm, a plurality of control behaviors for retrieving items by holding each item at a determined contact point using a suction device of the robotic arm;
evaluating the plurality of control behaviors for the robotic arm, by estimating a seal quality metric for each control behavior, based on a predicted deformed n-dimensional shape of the suction device and a n-dimensional shape associated with an item; and
refining the machine learning model, based on the estimated seal quality metrics.
16. The non-transitory computer-readable medium of claim 15 , wherein dynamically generating the plurality of synthetic machine learning model input data using the synthetic input generation engine comprises, for at least one of the plurality of synthetic machine learning model input data:
selecting a first item from a product catalog for a fulfillment center; and
retrieving, using a unique identifier corresponding to the first item, an n-dimensional model corresponding to the first item.
17. The non-transitory computer-readable medium of claim 16 , wherein dynamically generating the plurality of synthetic machine learning model input data using the synthetic input generation engine further comprises, for the at least one of the plurality of synthetic machine learning model input data:
dynamically generating a pose for the first item using the synthetic input generation engine;
determining a surface geometry for the first item, based on the generated pose and the retrieved n-dimensional model corresponding to the first item; and
selecting a candidate contact point on the surface geometry for the first item.
18. The non-transitory computer-readable medium of claim 17 , wherein estimating the seal quality metric for each control behavior comprises:
modelling deformation and contact between the suction device and the surface geometry for the first item at the respective candidate contact point, using at least one n-dimensional spring-lattice to model deformation and contact between the suction device and the surface geometry for the first item at the respective candidate contact point; and
calculating the seal quality metric, using the n-dimensional spring-lattice and a fluids model relating fluid force and leak pathways.
19. The non-transitory computer-readable medium of claim 18 , wherein estimating the seal quality metric for each control behavior further comprises
iteratively refining a n-dimensional shape of a first one of the at least one n-dimensional spring-lattice to minimize potential energy across the first n-dimensional spring-lattice, wherein a resulting n-dimensional shape of the first n-dimensional spring-lattice represents the n-dimensional shape of the suction device when suction is applied to the respective candidate contact point of the first item.
20. The non-transitory computer-readable medium of claim 19 , wherein estimating the seal quality metric for each control behavior further comprises
iteratively refining a second n-dimensional shape of a second one of the at least one n-dimensional spring-lattice to minimize potential energy across the second n-dimensional spring-lattice, wherein a resulting n-dimensional shape of the second n-dimensional spring-lattice represents the n-dimensional shape for the respective candidate contact point on the first item when suction is applied using the suction device.