The article also outlines the future plans of comma.ai. They aim to ship a model trained in a new architecture by October 2024 and eventually allow on-device Reinforcement Learning training. The team believes that their generic learning strategies can be applied beyond driving to all of robotics, with the goal of creating open-source robots that can perform household tasks.
Key takeaways:
- The autonomy team at comma.ai is focused on building open-source robotic agents, with their main product being an advanced driver-assistance system (ADAS).
- The team has achieved end-to-end lateral and longitudinal control in their driving models, and has also experimented with end-to-end navigation, although this feature was temporarily removed to focus on improving the longitudinal policy.
- Comma.ai is developing a new simulator using a generative model to improve the training of their driving models. They also plan to incorporate Reinforcement Learning (RL) to optimize their models and controllers.
- The methods used by comma.ai are not specific to driving and can be applied to other areas of robotics. The company envisions creating open-source robots for household tasks in the future.