Control system using autoencoder

1. A control system comprising:

a memory storing a sequence of sensor data received from one or more sensors;

one or more processors which process the sensor data to compute a sequence of derived sensor data values; and

an autoencoder configured to:

receive the sequence of derived sensor data values;

compute a forward prediction of the sequence of derived sensor data values in latent space; and

decode the forward prediction of the sequence of derived sensor data values in latent space to provide a forward prediction of the sequence of derived sensor data values, the autoencoder having been trained by imposing a relationship on positions of sensor data values encoded in the latent space of the autoencoder; and

at least one of the processors being configured to initiate control of an apparatus using the forward prediction of the sequence of derived sensor data values.

2. The control system of claim 1 wherein the relationship comprises a motion model.

3. The control system of claim 1 wherein the relationship comprises a velocity.

4. The control system of claim 3 wherein the relationship comprises a constant velocity.

5. The control system of claim 1 wherein the relationship comprises an acceleration.

6. The control system of claim 1 wherein the processor uses a parameterized generative model of the derived sensor data values, and wherein the parameters of the generative model are related to parameters of the derived sensor data values when encoded in the latent space.

7. The control system of claim 6 wherein the parameters of the generative model are selected to correspond with parameters of the derived sensor data values when encoded in the latent space.

8. The control system of claim 1 wherein the autoencoder has been trained using an objective function having at least one error term defined in the latent space.

9. The control system of claim 1 wherein the autoencoder has been trained using an objective function having an error term encouraging latent variables to have identity covariance.

10. The control system of claim 1 wherein the sequence of derived sensor data values comprises at least three consecutive values x1, x2, x3, and wherein the autoencoder has been trained using an objective function which comprises an error term which is related to a motion model of the at least three consecutive values in the latent space.

11. The control system of claim 1 wherein an objective function penalizes differences between inputs to the encoder and corresponding outputs of the decoder, and where the error term which is related the a motion model is computed as a norm of the at least three consecutive values in the latent space.

12. A computer-implemented method comprising:

receiving a sequence of derived sensor data values computed from a sequence of sensor data received from one or more sensors;

compute a forward prediction of the sequence of derived sensor data values in latent space

decode the forward prediction of the sequence of derived sensor data values in latent space to provide a forward prediction of the sequence of derived sensor data values, the autoencoder having been trained by imposing a relationship on positions of sensor data values encoded in the latent space of the autoencoder.

13. The method of claim 12 comprising initiating control of an apparatus using the forward prediction.

14. The method of claim 12 comprising training the autoencoder using an objective function having an error term defined in the latent space.

15. The method of claim 12 wherein the sequence of derived sensor data values comprises at least three consecutive values x1, x2, x3, and wherein the method comprises training the autoencoder using an objective function which penalizes differences between inputs to the encoder and corresponding outputs of the decoder, and which comprises an error term which is the square of double the encoding of x2 in the latent space, minus the encoding of x1 in the latent space, minus the encoding of x3 in the latent space.

16. A control system comprising:

a memory storing a sequence of sensor data received from one or more sensors, the sensor data depicting an object;

a pose detector which processes the sensor data to compute a sequence of values of pose parameters of the object by fitting a parameterized model to the sensor data;

an autoencoder configured to:

receive the sequence of pose parameter values;

compute a forward prediction of the sequence of derived sensor data values in latent space; and

decode the forward prediction of the sequence of derived sensor data values in latent space to provide a forward prediction of the sequence of pose parameter values, the autoencoder having been trained by imposing a relationship on positions of the pose parameter values encoded in the latent space of the autoencoder; and

a processor configured to initiate control of an apparatus using the forward prediction of the sequence of derived sensor data values.

17. The control system of claim 16 wherein the parameterized model is a 3D model of the object and wherein the parameters of the model are selected according to the encodings of the pose parameter values in the latent space.

18. The control system of claim 16 wherein the pose detector fits the parameterized model to the sensor data by computing an optimization and wherein the optimization omits a prior defining how the values of the parameters change over time.

19. The control system of claim 16 wherein the processor initiates control of a human-computer interface using the forward prediction.

20. The control system of claim 16 wherein the relationship comprises a motion model.