Physics informed neural network for learning non-Euclidean dynamics in electro-mechanical systems for synthesizing energy-based controllers

1. A system of neural networks for synthesizing a controller, comprising:

a processor; and

a non-transitory memory having stored thereon modules executed by the processor, the modules comprising:

a feeder neural network trained to estimate an ordinary differential equation (ODE) from time series training data (X) of a trajectory of a robot having a combination of embedded angular data and generalized coordinate and momentum data, configured to learn dynamics of a physical system that includes the robot by encoding a generalization of a Hamiltonian representation of the dynamics using a constant external control term (u) for driving parameter learning in the feeder neural network; and

a neural ordinary differential equation (ODE) solver configured to receive the estimate of the ODE from the feeder neural network, and to synthesize a controller to control the system to track a reference configuration, wherein the controller is configured to control the robot based on the reference configuration that is tracked.

2. The system of claim 1 , wherein the feeder neural network comprises function approximator neural networks, comprising:

a first neural network to parameterize an inverse mass matrix for a first term in the Hamiltonian representation;

a second neural network to parameterize a potential energy matrix for a second term in the Hamiltonian representation; and

a third neural network to parameterize an input matrix as a scaler term for the external control term (u).

3. The system of claim 1 , wherein feeder neural network includes a partial derivative operation on the Hamiltonian representation solved by automatic differentiation.

4. The system of claim 1 , wherein training data is generated based on whether translational degrees of freedom and rotational degrees of freedom are present in the dynamical system and whether generalized momenta data is available.

5. The system of claim 1 , wherein the feeder neural network architecture is defined based on whether translational degrees of freedom and rotational degrees of freedom are present in the dynamical system and whether generalized momenta data is available.

6. The system of claim 1 , wherein the feeder neural network infers energy parameters from the learned dynamics and incorporates energy shaping by deriving a potential energy shaping function such that a desired Hamiltonian has a minimum at a desired reference configuration to ensure lowest energy in the system at the reference configuration.

7. A method for synthesizing a controller, comprising:

using a feeder neural network trained to estimate an ordinary differential equation (ODE) from time series training data (X) of a trajectory of a robot having a combination of embedded angular data and generalized coordinate and momentum data;

learning dynamics of a physical system that includes the robot by encoding a generalization of a Hamiltonian representation of the dynamics using a constant external control term (u) for driving parameter learning in the feeder neural network;

receiving the estimate of the ODE from the feeder neural network; and

synthesizing a controller to control the system to track a reference configuration; and

controlling the robot based on the reference configuration that is tracked.

8. The method of claim 7 , wherein the feeder neural network comprises function approximator neural networks, further comprising:

using a first neural network to parameterize an inverse mass matrix for a first term in the Hamiltonian representation;

using a second neural network to parameterize a potential energy matrix for a second term in the Hamiltonian representation; and

using a third neural network to parameterize an input matrix as a scaler term for the external control term (u).

9. The method of claim 7 , wherein feeder neural network includes a partial derivative operation on the Hamiltonian representation solved by automatic differentiation.

10. The method of claim 7 , further comprising:

generating training data based on whether translational degrees of freedom and rotational degrees of freedom are present in the dynamical system and whether generalized momenta data is available.

11. The method of claim 7 , further comprising:

defining the feeder neural network architecture based on whether translational degrees of freedom and rotational degrees of freedom are present in the dynamical system and whether generalized momenta data is available.

12. The method of claim 7 , further comprising:

inferring energy parameters from the learned dynamics and incorporating energy shaping by deriving a potential energy shaping function such that a desired Hamiltonian has a minimum at a desired reference configuration to ensure lowest energy in the system at the reference configuration.