The Mule: the forbidden love-child issuing from a male donkey (a jack) and a mare. Sure footed. Even tempered. Ok, maybe your friends laugh at you from their fancy horses, their thoroughbreds, their imported Arabians; but we'll see who gets the last laugh!
muleAI is inspired by the DonkeyCar project. We decided not to fork, but to rewrite.
muleAI is a systematic re-implementation of some core functionality with some priorities in mind:
- Simplicity and consistency in modular design
- Clean, well-structured implementation conforming to standard software design principles
muleAI is a lightweight python library that facilitates research and development in autonomous mobility at RC-scale.
muleAI is a foundation for further experiments in mobility, autonomous hardware, embedded AI, Internet Of Things, ...
UPDATE? Tensor flow 1.8 (includes keras as tf.keras
)
- Extended configuration file, YAML format
- As much as possible is exposed to configuration, allowing rapid changing of parameters during racing days
- Command line interface exposed using click
- States are saved using a timestamp, the
time.time() * 1000
(Unix standard, number of milliseconds since 1970)- Allows for fast timestep analysis, strict ordering and alignment of states
- Modular part classes inherit from Abstract Base Class
- Enforce proper interface for all new parts
- Include default behaviors such as class strings
- Extensive logging messages throughout the project, for faster debugging
- New adjustment method for PS3 controller, DPad selects a value, and up/down to change the value allowing performance changes as the car is driving
- D-pad left/right on the PS3 controller iterates over adjustment settings
- D-pad up/down on the PS3 controller adjusts that value by SHIFT amount
- Currently able to adjust max forward/reverse throttle and steering
- Images are saved directly to numpy arrays, timestamped, and zipped for fast transfer to training
- No support for any installation or setup method - project is run directly from the git directory
- Linux and Mac OS are tested, not Windows
- Training of models is in a separate module