Added dual policy mode to toggle default locomotion policy#73
Conversation
|
|
||
| from holosoma_inference.config.config_types.task import TaskConfig | ||
|
|
||
| _MODELS_DIR = Path(__file__).parent.parent.parent / "models" |
There was a problem hiding this comment.
could you add a comment clarifying where this should point to? Seems like src/holosoma_inference/holosoma_inference/models
|
|
||
| def main(annotated_config=None): | ||
| """Main entry point. Extensions can pass their own AnnotatedInferenceConfig.""" | ||
| import argparse |
There was a problem hiding this comment.
why not import at the top of the file?
tomasz-lewicki
left a comment
tomasz-lewicki
left a comment
There was a problem hiding this comment.
LGTM aside from the config. I think it should be pretty easy to skip all of this argparse logic, and let tyro handle this.
| restore_terminal_settings() | ||
|
|
||
|
|
||
| def _split_secondary_args(argv: list[str]) -> tuple[list[str], list[str]]: |
There was a problem hiding this comment.
Doesn't tyro handle this out-of-the box once you've defined TaskConfig.secondary? Maybe I'm missing something, but I think you can do:
run_policy.py --task.secondary.task.model-path weights.onnxWe can also just have a secondary optional InferenceConfig.
run_policy.py --task.model-path weights_dance.onnx --fallback_task.model-path weights_loco.onnx| id(self.secondary): self.secondary.handle_keyboard_button, | ||
| } | ||
|
|
||
| def patched_joy(cur_key): |
There was a problem hiding this comment.
Hopefully we refactor joystick/input handling soon so doing simple things doesn't have to be so convoluted.
asetapen
left a comment
asetapen
left a comment
There was a problem hiding this comment.
Very nice improvements and dual mode will be extremely helpful for testing.
I'm not too familiar with tyro so can't comment on the config issue.
5 participants
This PR adds as a default inference config the option to switch to a robust locomotion policy by pressing X on the joystick (or x on keyboard), which is convenient for the robot to recover while testing experimental policies. The implementation is even more general than that, as the way it works is by introducing a dual mode setup that allows switching between any two policies at runtime.