Comma's Openpilot 0.11 Launches First Robotics Agent Fully Trained in Learned Simulation
Key Takeaways
- ▸Openpilot 0.11 is the first shipping real-world robotics agent trained entirely in a learned simulation, marking a paradigm shift from hand-coded simulators to neural network-based training environments
- ▸The new commaVQ world model generates significantly better synthetic driving video quality and temporal consistency, enabling training on previously difficult scenarios like longitudinal speed control
- ▸User adoption data shows Experimental mode (end-to-end longitudinal policy) achieving its highest usage rates yet with v0.11, indicating real-world performance improvements in speed convergence and obstacle reactivity
Summary
Comma has released Openpilot 0.11, marking a major milestone in autonomous driving by shipping the first real-world robotics agent fully trained in a learned simulation environment. The new version features a driving model trained entirely on synthetic data generated by commaVQ, a neural network-based world model, replacing the hand-coded simulators and motion planning controllers used in previous versions. This shift from classical simulation to learned simulation enables the model to handle a much wider range of driving scenarios, particularly improving speed convergence on highways and reactivity around parked cars.
The breakthrough represents the culmination of a decade-long research initiative that began in 2016 with Comma's foundational paper "Learning a Driving Simulator." The company progressively moved toward end-to-end learning, releasing transformer-based world models in 2023 and planning models trained with world models in 2025, before achieving full learned-simulation training in this release. Beyond the advanced driving model, Openpilot 0.11 also delivers a 77% reduction in idle power draw and improved user interface polish for Comma Four devices.
- The release includes major efficiency improvements with 77% reduction in idle power consumption and enhanced interface design for Comma Four hardware
Editorial Opinion
This represents a genuine inflection point in autonomous driving development. Training a real-world robotics agent entirely in learned simulation—rather than relying on hand-coded physics engines—addresses fundamental limitations in classical simulation and opens the door to scaling driving capabilities in ways previously constrained by simulator realism. The fact that users are actively preferring the experimental end-to-end mode validates the practical benefits of this approach, though the long-term safety implications of learned simulators in safety-critical domains will require sustained attention.
