K-Infer

kinfer is a package for doing inference on K-Scale models, by providing a standard interface for trained policies. This lets you take a trained policy, compile it to a kinfer model binary, and run the same binary in simulation and on a real robot.

Each kinfer model has three parts:

1. An init function, which initializes the model carry
2. A step function, which processes one computation step
3. Some additional metadata telling downstream implementations how to use the model input and outputs

The control loop for a kinfer model can be thought of as follows:

carry = init() # Onnx graph
while True:
  model_input = get_model_input() # Implemented by provider
  model_output, carry = step(model_input, carry) # Onnx graph
  do_model_output(model_output) # Implemented by provider
  sleep()

This provides a common interface for running models in simulation and on the real robot, making sim2real transfer straightforward.