Tesla recently released the full self-driving (FSD) Beta on highways. This new system is designed to replace the legacy highway stack and unifies the vision and planning stack both on and off-highway. The legacy highway stack relied on multiple single-camera and single-frame networks which were designed to handle simple lane-specific maneuvers. However, the FSD Beta’s multi-camera video networks and next-gen planner allows for more complex agent interactions with less reliance on lanes, which means more intelligent behaviors, smoother control, and better decision-making.
Tesla has also made a number of improvements that will benefit drivers on highways. For example, recall for close-by cut-in cases has increased by 15%, particularly for large trucks and high-yaw rate scenarios. This was achieved by adding additional 30k auto-labeled clips mined from the fleet. Additionally, dedicated speed control for cut-in objects has been expanded and tuned.
To improve the position of the ego on wide lanes, Tesla has also implemented a system that biases in the direction of the upcoming turn to allow other cars to maneuver around the ego. This is especially useful for scenarios with high curvature or large trucks, as the ego will now be able to offset in lanes to maintain safe distances to other vehicles on the road and increase comfort.
Furthermore, Tesla has improved the behavior for path blockage lane changes in dense traffic, allowing for more headway in blocked lanes to hedge for possible cars in dense traffic. To make lane changes in dense traffic scenarios smoother, Tesla has also allowed for higher acceleration during the alignment phase, resulting in more natural gap selection to overtake adjacent lane vehicles very close to the ego.
Tesla has also improved the detection consistency between lanes, lines, and road edge predictions, making turns smoother. Additionally, the accuracy of detecting other vehicles’ moving semantics has been improved, with a 23% increase in precision for cases where other vehicles transition to driving and a 12% reduction in error for cases where Autopilot incorrectly detects its lead vehicle as parked. This was achieved by increasing the video context in the network, adding more data on these scenarios, and increasing the loss penalty for control-relevant vehicles.
Finally, Tesla has extended the maximum trajectory optimization horizon, which results in smoother control for high curvature roads and far away vehicles when driving at highway speeds. The company has also improved the driving behavior next to rows of parked cars in narrow lanes, preferring to offset and stay within lane instead of unnecessarily lane changing away or slowing down. False offsetting around objects in wide lanes and near intersections has also been reduced by improving object kinematics modeling in low-speed scenarios.
In conclusion, the FSD Beta on highways has been designed to bring numerous improvements to the driving experience, from making turns smoother to improving position of the ego on wide lanes and reducing false offsetting around objects. With the addition of the multi-camera video networks and next-gen planner, the system is allowing for more complex agent interactions with less reliance on lanes, making way for more intelligent behaviors, smoother control, and better decision-making.