InterSim Interactive Traffic Simulation via Explicit Relation Modeling Qiao Sun1 Xin Huang2 Brian C. Williams2 and Hang Zhao1 Abstract Interactive traffic simulation is crucial to au-

2025-05-05 0 0 1.63MB 8 页 10玖币
侵权投诉
InterSim: Interactive Traffic Simulation via Explicit Relation Modeling
Qiao Sun1, Xin Huang2, Brian C. Williams2, and Hang Zhao1∗
Abstract— Interactive traffic simulation is crucial to au-
tonomous driving systems by enabling testing for planners in a
more scalable and safe way compared to real-world road testing.
Existing approaches learn an agent model from large-scale
driving data to simulate realistic traffic scenarios, yet it remains
an open question to produce consistent and diverse multi-
agent interactive behaviors in crowded scenes. In this work,
we present InterSim, an interactive traffic simulator for testing
autonomous driving planners. Given a test plan trajectory
from the ego agent, InterSim reasons about the interaction
relations between the agents in the scene and generates realistic
trajectories for each environment agent that are consistent with
the relations. We train and validate our model on a large-
scale interactive driving dataset. Experiment results show that
InterSim achieves better simulation realism and reactivity in
two simulation tasks compared to a state-of-the-art learning-
based traffic simulator.
I. INTRODUCTION
With the recent development of autonomous driving tech-
nologies, traffic simulation has played an important role
in enabling testing the planner system on a large scale.
Compared to real-world road testing, simulation offers a
more time and resource efficient alternative by reconstructing
rare but important traffic scenarios. More importantly, it
allows simulating risky scenarios that are usually difficult
to obtain in real-world driving.
Traditional simulators often rely on static log replay that
simulates the agent behavior based on what happened in
the collected data. It fails to produce reactive behavior of
environment agents when the ego plan diverges from the
original log and thus becomes less useful in interactive
scenarios. On the other hand, heuristic-based models, such
as the intelligent driver model (IDM) [1]–[3], produce more
reactive behaviors in response to diverging ego plans, but
they are limited to following predefined trajectories and have
difficulties in producing diverse scenarios.
Recent advances in machine learning have enabled real-
istic and diverse agent simulation by training agent models
from realistic driving data. They demonstrated great potential
in supporting closed loop planner evaluation. For instance,
[4] proposes to generate realistic driving episodes by leverag-
ing a probabilistic prediction model given traffic observations
and environmental context; [5] infers future agent states as
both discrete intent and continuous controls conditioned on
past observations over all agents. On the other hand, they
focus on simulating trajectories of individual agents without
reasoning about their future interactions, which could lead
1IIIS, Tsinghua University alan.qiao.sun@gmail.com
2CSAIL, Massachusetts Institute of Technology
∗Corresponding at hangzhao@mail.tsinghua.edu.cn
Fig. 1: A motivating example of InterSim. Given an updated
ego plan in red, InterSim first identifies the relevant agent
whose future trajectories may collide with the ego plan, and
infers the interacting relations between the agents, such as
the environment agent is yielding to the ego agent. It then
predicts consistent and reactive trajectories for the relevant
agent given the inferred relations. The simulator simulates
the next step of all agents based on the predictions and waits
for the new plan.
to colliding trajectories in dense traffic. To overcome this
challenge, [6] adds a task loss to penalize collisions and [7]
proposes a feasibility check on the generated trajectories
to filter out colliding trajectories. Instead of requiring a
hand-crafted loss or an ad-hoc filter, [8] offers simulation
consistency by rolling outs joint trajectories over all the
agents in a scene through an implicit latent variable learned
by a conditional variational autoencoder; however, such gen-
erative models rely on probabilistic sampling and suffer from
producing rare or dangerous scenarios, which are crucial to
testing autonomous driving planners.
In this work, we propose InterSim, an interactive traffic
simulation model that reasons about explicit interaction re-
lations when rolling out future trajectories for all agents in
a scene. As shown in Fig. 1, given an updated ego plan
in red, InterSim first identifies the relevant agents whose
future trajectory may collide with the new ego plan. Next, it
infers the interaction relations between the ego agent and the
relevant environment agents, and predicts reactive trajectories
for the relevant agents based on the relation. Finally, it
follows the predictions to simulate one step ahead for the
environment agents and repeat the same procedure until the
end of the simulation episode.
Compared to existing learning-based models, InterSim
offers a few advantages by reasoning about the explicit
interactive relations among the relevant traffic agents. First,
the relations guide the trajectory simulator to produce con-
sistent trajectories of multiple agents in complex interactive
scenarios and offer better interpretability. Second, one can
use it to manipulate an interactive scenario by specifying the
arXiv:2210.14413v1 [cs.RO] 26 Oct 2022
interaction relation between agents, which is a non-trivial
task for generative models that may require a large number
of samples. Third, it affords better efficiency by identifying
only the relevant agents influenced by the ego plans and
modifying their future trajectories in simulation as opposed
to all agents.
Our contribution is as follows:
•We propose a learning-based simulation model,
InterSim, that rolls out realistic and consistent future
trajectories of multiple traffic agents based on explicit
interaction relations.
•We leverage a relation predictor to infer interaction
relations for better interpretability and simulation ef-
ficiency, and show how our simulator can be used to
manipulate different interaction situations by specifying
the relations.
•We train and evaluate our model on the Waymo Open
Motion Dataset, a publicly available real-world driving
benchmark, and demonstrate its advantage compared to
a state-of-the-art baseline in two simulation tasks.
II. RELATED WORK
In this section, we discuss relevant literature in three as-
pects: traffic simulation, behavior prediction, and interaction
modeling.
A. Traffic Simulation
Traffic simulation is an important task for intelligent
transportation systems, allowing for training and evaluating
driving models in a more scalable and safe way. Existing traf-
fic simulators render high-fidelity driving environments in the
context of racing [9] and urban driving [10], [11]. However,
they often simulate agent behaviors through heuristic-based
models that fail to cover diverse scenarios or interactions.
Recently, learning-based models have demonstrated great
success in simulating realistic and reactive agent behaviors
by learning driving patterns from real-world driving data. For
instance, [4] trains a deep neural network through a rasterized
representation derived from driving logs to simulate future
agent trajectories; [5] infers future agent intent and control
inputs to model stochastic traffic dynamics. While such
methods consider the past trajectories of all the agents at
once, they assume independence of future trajectory rollouts
that may lead to inconsistent or colliding trajectories between
simulated agents in interactive scenarios.
In order to improve simulation consistency over multiple
interacting agents, [6] leverages a collision loss and [7]
proposes a rule-based fall-back layer to discourage or avoid
collisions. While such works often require hand-crafted
losses or post-processing filters, [8] proposes a multi-agent
behavior model that simulates joint agent behaviors directly
through an implicit latent variable that governs the agent
interactions. Compared to existing models, we propose a
relation-aware simulator that simulates diverse and realistic
interactive behaviors in a more straightforward and efficient
way by explicitly modeling interacting relations.
B. Behavior Prediction
Behavior prediction offers a natural solution to simulate
agent behaviors through the predicted trajectories given the
environmental context. Recent models prove great success in
improving prediction accuracy, by learning agent dynamics
and environmental context represented either as a vector
representation [12], [13] or a rasterized image [14], [15].
Due to uncertainty in human intent, the future trajectories
are multi-modal. To handle the multi-modality and improve
prediction coverage, a family of models are proposed to first
predict high-level intent, such as goal targets [16]–[18], lanes
to follow [19], [20], maneuvers [21]–[23], and linguistic
descriptions [24], before predicting low-level trajectories that
are conditioned on the intent.
In this work, we take advantage of the goal-conditioned
models in the behavior prediction literature to simulate
realistic agent trajectories given the environmental context
and agent intent.
C. Interaction Modeling
Modeling interaction is an important task in motion pre-
diction and simulation when reasoning about multi-agent be-
haviors. While many existing approaches [?], [25]–[27] rely
on implicit latent variables to model interactions, we focus on
modeling and predicting explicit interaction relations in this
work for better interpretability. These explicit relations allow
us to produce and manipulate different types of interactive
scenarios.
In this work, we follow [28]–[30] that define agent re-
lations based on the pass and yield relationship and predict
the relationship as a classification problem through a separate
learning model. The predicted relations are useful in guid-
ing the motion predictor to generate consistent trajectories
among multiple agents, as shown by [30].
When there exist potential conflicts between a novel ego
plan and the simulated trajectories of environment agents
given the predicted relations, we adopt conflict resolution
techniques that are widely used in planning [31], search [32],
and ordering [33].
III. PROBLEM FORMULATION
We formulate the problem of learning realistic interactive
behaviors for traffic simulation following [8]. Given map
states Mand the observed states Sof Ntraffic agents in a
scene, the goal is to roll out the future states of all agents Y
up to a finite horizon T.
Due to the computational complexity and memory con-
straint in simulating joint behaviors over all traffic agents in
the scene, our model focuses on simulating agent behaviors
that are relevant to the ego plan, as the irrelevant agent
behaviors are often ignored by the ego planner. For an
irrelevant agent whose future trajectory stays the same given
a new ego plan at the next step, our simulator can simply
roll out its future trajectory from the data.
One key consideration in our problem is to faithfully
follow the agent’s origin intent as much as possible. We
define such intent based on the goal location collected from
摘要：

InterSim:InteractiveTrafcSimulationviaExplicitRelationModelingQiaoSun1,XinHuang2,BrianC.Williams2,andHangZhao1Abstract—Interactivetrafcsimulationiscrucialtoau-tonomousdrivingsystemsbyenablingtestingforplannersinamorescalableandsafewaycomparedtoreal-worldroadtesting.Existingapproacheslearnanagentm...

展开>> 收起<<
InterSim Interactive Traffic Simulation via Explicit Relation Modeling Qiao Sun1 Xin Huang2 Brian C. Williams2 and Hang Zhao1 Abstract Interactive traffic simulation is crucial to au-.pdf

共8页,预览2页

还剩页未读， 继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

相关推荐

更多
立即下载
分类：图书资源 价格：10玖币 属性：8 页 大小：1.63MB 格式：PDF 时间：2025-05-05

开通VIP享超值会员特权

  • 多端同步记录
  • 高速下载文档
  • 免费文档工具
  • 分享文档赚钱
  • 每日登录抽奖
  • 优质衍生服务

作者详情

相关内容

更多

热门标签

人际关系 动力学连接体 力的合成 配电装置 高考理综 全宋诗作者索引 公务员考试
/ 8
评分 收藏
分享
立即下载
客服
关注