1. Introduction to nuplan:
1.1 Overview of the content and framework of the nuplan dataset
It is the world's first open source data set (released in 2021) for autonomous driving planning method tests. It collected approximately 1,300 hours of driving data in four cities: Boston, Pittsburgh, Las Vegas and Singapore.
Although ML-based motion planners are increasingly available, the lack of established datasets, simulation frameworks, and evaluation metrics also limits progress in this field. Existing datasets ( Argoverse , Lyft , Waymo ) mainly focus on short-term motion prediction of other agents, rather than long-term planning of the ego vehicle. This leads to the fact that most of the previous methods use open-loop evaluation based on L2 indicators, which is not suitable for rational evaluation of long-term planning. In order to solve the above problems, in addition to providing large-scale real vehicle data sets, nuplan also provides a learning-based training framework to develop machine learning-based planners, lightweight closed-loop simulators, motion planning evaluation indicators and visualization tools . The figure below is its training framework.
1.2 Overview of nuplan emulator
Nuplan's open-source simulation framework supports both closed-loop and open-loop simulations. Closed-loop means that the ego vehicle and other vehicles can deviate from the state information recorded in the original dataset . In addition to the baseline implementation of the planner, nuplan also provides baseline implementations for traditional Intelligent Driver Model (Intelligent Driver Model, IDM) and learning-based (ML) agents . In order to realistically simulate traffic flow, nuplan has also developed a traffic light state inference system , which infers the state of traffic lights from the observed vehicle motion in the scene. Finally, nuplan also mines hand-crafted scenarios (e.g., lane changes, pedestrian-vehicle interactions) and evaluates them on both generic and scene-specific metrics.
2. Nuplan planning framework
nuPlan has an open source toolkit for developing learning (ML) based planners. The nuplan planning framework is generally facilitated by the following four steps, namely: model training , model simulation , performance evaluation and result visualization .
2.1 Planning training framework
The nuplan training learning-based (ML) planner is based on the PyTorch framework , using some inputs, such as self-vehicle information , other agent information , and static/dynamic map information . Based on this framework, it can be easily modified and extended to incorporate new research ideas.
2.2 Planning Simulation Framework
The core part of the nuplan framework is scene simulation. The simulator fully matches the dataset and integrates modules such as observation model , ego planner , agent model , motion controller , evaluation metrics and visualization . The emulator also provides multiple options and basic examples for each module, allowing users to easily plug in replacements based on individual needs. This will allow the research community to focus methodologically on a single framework, rather than using different frameworks for each dataset and planning method as has been the case previously.
The figure below shows the overall framework of the nuplan emulator.
2.3 Evaluation criteria
The evaluation indicators are divided into 5 categories:
- Traffic rule violations are used to measure compliance with general traffic rules, taking into account the collision rate of other vehicles, the probability of trajectory deviation from the road, the time gap with the vehicle in front, the time of collision, and the relative speed when overtaking.
- Similarity to human driving was used to quantify handling satisfaction compared to humans, such as longitudinal speed error, longitudinal stop position error, and lateral position error. Furthermore, the resulting tics/accelerations were compared to human-level tics/accelerations.
- Vehicle dynamics are used to evaluate driver comfort and trajectory feasibility. Occupant comfort is measured by jerk, acceleration, steering rate and vehicle oscillations. Feasibility is measured by violating predetermined constraints against the same criteria.
- The achievement of the goal evaluates the route to the goal location on the map by the L2 distance
-
Scenario-specific evaluations are for specific scenarios. For example, for a vehicle lane change, the time to collision and the time gap to the front/rear people in the target lane are evaluated. For vehicle interactions with pedestrians and cyclists, the relative velocity of passage is quantified while distinguishing their locations.
2.4 Visualization
For researchers to rapidly iterate on new ML planning models, it is critical to be able to carefully analyze the performance of the models. To this end, nuplan has developed the 'nuBoard' with an interactive bokeh-based simulation viewer that renders the underlying semantic map , ego vehicle information , other vehicle information , and traffic light status . In addition to scene simulations, nuplan also plots statistics about the performance of the planner.
3. Summarize the characteristics of nuplan
1. Abundant data, providing rich information of self-vehicle, other-vehicle, traffic status and sensors in four cities.
2. Integrate the automatic driving planning framework based on machine learning, enabling rapid testing and verification of academic methods.
3. Integrated simulation function, the self-vehicle and other vehicles in the data set can deviate from the real location information collected, and realize the closed-loop simulation test
4. Reference content
motional/nuplan-devkit: The devkit of the nuPlan dataset. (github.com)