Sensorized robotic gripping device

1. A method comprising:

controlling one or more digits of a robotic gripping device to attempt to grasp an object;

receiving, from a first non-contact sensor positioned outside of the one or more digits of the robotic gripping device, first sensor data based on a first sensing modality, wherein the first sensor data is collected after the attempt to grasp the object;

receiving, from a second non-contact sensor positioned outside of the one or more digits of the robotic gripping device, second sensor data based on a second sensing modality different from the first sensing modality, wherein the second sensor data is collected after the attempt to grasp the object; and

determining, using a trained machine learning model that takes as input the first sensor data and the second sensor data and outputs a decision of whether the robotic gripping device is currently holding the object after the attempt to grasp the object, whether the robotic gripping device is currently holding the object after the attempt to grasp the object.

2. The method of claim 1 , wherein the trained machine learning model has been trained based on past grasp attempts by the robotic gripping device.

3. The method of claim 1 , wherein the trained machine learning model has been trained based on past grasp attempts by a similar robotic gripping device.

4. The method of claim 1 , wherein the trained machine learning model has been trained based on simulated grasping events.

5. The method of claim 1 , wherein the trained machine learning model comprises a convolutional neural network (CNN).

6. The method of claim 1 , wherein the first non-contact sensor comprises a time-of-flight sensor, wherein the first sensor data comprises time-of-flight distance data from the time-of-flight sensor.

7. The method of claim 6 , wherein the trained machine learning model is a linear kernel state vector machine (SVM) that takes as input the time-of-flight distance data.

8. The method of claim 1 , wherein the first non-contact sensor is an infrared camera, wherein the first sensor data comprises grayscale image data, wherein the second non-contact sensor is a time-of-flight sensor, and wherein the second sensor data comprises time-of-flight distance data.

9. The method of claim 8 , wherein the trained machine learning model takes as further input third sensor data, wherein the third sensor data comprises reflectance data from the time-of-flight sensor.

10. The method of claim 8 , wherein the trained machine learning model is an SVM that takes as input a single image from the infrared camera, a single distance measurement from the time-of-flight sensor, and a single reflectance measurement from the time-of-flight sensor.

11. The method of claim 1 , further comprising receiving, from an additional sensor on the robotic gripping device, additional sensor data, wherein the trained machine learning model takes as further input the additional sensor data.

12. The method of claim 1 , further comprising controlling the robotic gripping device to attempt to grasp the object a second time when the attempt to grasp the object results in a grasp failure.

13. The method of claim 1 , wherein the robotic gripping device comprises a plurality of underactuated digits, wherein each underactuated digit has less control inputs than degrees of freedom.

14. The method of claim 1 , wherein the first non-contact sensor is a one-dimensional (1D) time-of-flight sensor configured to generate a time-of-flight distance measurement indicative of distance to a nearest object in a direction extending between a plurality of digits of the robotic gripping device.

15. A robot comprising:

a robotic gripping device comprising one or more digits;

a first non-contact sensor positioned outside of the one or more digits of the robotic gripping device;

a second non-contact sensor positioned outside of the one or more digits of the robotic gripping device; and

a control system configured to:

control the robotic gripping device to attempt to grasp an object;

receive, from a first non-contact sensor, first sensor data based on a first sensing modality, wherein the first sensor data is collected after the attempt to grasp the object;

receive, from a second non-contact sensor, second sensor data based on a second sensing modality different from the first sensing modality, wherein the second sensor data is collected after the attempt to grasp the object; and

determine, using a trained machine learning model that takes as input the first sensor data and the second sensor data and outputs a decision of whether the robotic gripping device is currently holding the object after the attempt to grasp the object, whether the robotic gripping device is currently holding the object after the attempt to grasp the object.

16. The robot of claim 15 , where the first non-contact sensor is a camera.

17. A method comprising:

controlling one or more digits of a robotic gripping device to attempt to grasp an object;

receiving, from a first non-contact sensor positioned outside of the one or more digits of the robotic gripping device, first sensor data based on a first sensing modality, wherein the first sensor data is collected after the attempt to grasp the object;

receiving, from a second non-contact sensor positioned outside of the one or more digits of the robotic gripping device, second sensor data based on a second sensing modality different from the first sensing modality, wherein the second sensor data is collected after the attempt to grasp the object; and

training, based on the first sensor data, the second sensor data, and a result of the attempt to grasp the object, a machine learning model configured to take input of the first sensing modality and the second sensing modality and output a decision of whether the robotic gripping device is currently holding a given object after the attempt to grasp the given object.

18. The method of claim 17 , further comprising training the machine learning model on past grasp attempts by a similar robotic gripping device.

19. The method of claim 17 , further comprising training the machine learning model based on simulated grasping events.

20. The method of claim 17 , wherein the machine learning model comprises a convolutional neural network (CNN).