Hybrid computing achitectures with specialized processors to encode/decode latent representations for controlling dynamic mechanical systems

1. A dynamic mechanical system comprising:

at least one sensor of a first sensor type, each sensor of the first sensor type to sense a first property of the dynamic mechanical system or of an environment of the dynamic mechanical system;

at least a first hardware machine-learning accelerator coupled to a respective output of each of the at least one sensor of the first sensor type, the at least a first hardware machine-learning accelerator comprising an integrated circuit that hardcodes into circuitry a processing of the output of each of the at least one sensor of the first sensor type;

at least one sensor of a second sensor type, each sensor of the second sensor type to sense a second property of the dynamic mechanical system or of an environment of the dynamic mechanical system, wherein the second sensor type is different from the first sensor type and the second property is different from the first property;

at least a second hardware machine-learning accelerator coupled to a respective output of each of the at least one sensor of the second sensor type, the at least a second hardware machine-learning accelerator comprising an integrated circuit that hardcodes into circuitry a processing of the output of each of the at least one sensor of the second sensor type; and

a processor coupled to respective outputs of the at least a first hardware machine-learning accelerator and the at least a second hardware machine-learning accelerator, the processor to execute a control model for controlling the dynamic mechanical system based at least in part on data received by the processor from the at least a first hardware machine-learning accelerator and the at least a second hardware machine-learning accelerator.

2. The dynamic mechanical system of claim 1 wherein the dynamic mechanical systems includes at least one robot that carries the at least one sensor of the first sensor type, the at least a first hardware machine-learning accelerator, the at least one sensor of the second sensor type, and the at least a second hardware machine-learning accelerator.

3. The dynamic mechanical system of claim 1 wherein the at least a first hardware machine-learning accelerator is coupled to the at least one sensor of the first sensor type by a first high-speed bus.

4. The dynamic mechanical system of claim 3 wherein the at least a second hardware machine-learning accelerator is coupled to the at least one sensor of the second sensor type by a second high-speed bus, the second high-speed bus different form the first high-speed bus.

5. The dynamic mechanical system of claim 1 wherein the at least a first hardware machine-learning accelerator and the at least one sensor of the first sensor type are both carried on a first printed circuit board.

6. The dynamic mechanical system of claim 5 wherein the at least a second hardware machine-learning accelerator and the at least one sensor of the second sensor type are both carried on a second printed circuit board, the second printed circuit board different form the first printed circuit board.

7. The dynamic mechanical system of claim 1 , further comprising:

at least one encoder that mediates coupling between the at least a first hardware machine-learning accelerator and the processor and between the at least a second hardware machine-learning accelerator and the processor.

8. The dynamic mechanical system of claim 1 wherein the at least a first hardware machine-learning accelerator reduces a dimensionality of data received from the at least one sensor of the first sensor type, and wherein the at least a second hardware machine-learning accelerator reduces a dimensionality of data received from the at least one sensor of the second sensor type.

9. The dynamic mechanical system of claim 1 wherein the processing of the output of each of the at least one sensor of the first sensor type that is hardcoded in circuitry of the first hardware machine-learning accelerator includes a first feature extractor to extract at least one feature about the first property of the dynamic mechanical system or of an environment of the dynamic mechanical system.

10. The dynamic mechanical system of claim 9 wherein the processing of the output of each of the at least one sensor of the second sensor type that is hardcoded in circuitry of the second hardware machine-learning accelerator includes a second feature extractor to extract at least one feature about the second property of the dynamic mechanical system or of an environment of the dynamic mechanical system, the second feature extractor different form the first feature extractor.

11. The dynamic mechanical system of claim 1 wherein the at least one sensor of the first sensor type includes an image sensor or camera.

12. The dynamic mechanical system of claim 11 wherein the at least one sensor of the second sensor type includes an inertial measurement unit.

13. The dynamic mechanical system of claim 11 wherein the at least one sensor of the second sensor type includes a tactile sensor.

14. The dynamic mechanical system of claim 11 wherein the at least one sensor of the second sensor type includes a force sensor.

15. The dynamic mechanical system of claim 11 wherein the at least one sensor of the second sensor type includes a position sensor.

16. The dynamic mechanical system of claim 1 wherein the at least one sensor of the first sensor type includes a position sensor and the at least one sensor of the second sensor type includes an inertial measurement unit.

17. The dynamic mechanical system of claim 16 wherein the at least one sensor of the first sensor type includes a position sensor and the at least one sensor of the second sensor type includes a tactile sensor.

18. The dynamic mechanical system of claim 1 wherein the at least one sensor of the first sensor type includes a position sensor and the at least one sensor of the second sensor type includes a force sensor.