Waymax: An Accelerated, Data-Driven Simulator for Large-Scale Autonomous Driving Research

May combine two claims.

Waymax provides features including pedestrians, cyclists, traffic lights, and inferred routes for goal-conditioned policy and progress metrics.

“Waymax simulates challenging obstacles present in urban driving, such as pedestrians and cyclists, and provides high-level route information for the ego vehicle to follow.”

• Also stated in: § 2 Related Work · plain

  “We provide a full set of features including pedestrians, cyclists, and traffic lights available in WOMD, and the inferred routes for goal-conditioned policy and progress metrics.”

• Also stated in: § 3 Simulator Features · plain

  “Waymax includes a suite of common metrics described in Sec.3.4, and several options for modeling the behavior of dynamic objects (vehicles and pedestrians) in the scene, outlined in Sec.3.5.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Waymax provides a set of metrics to evaluate the ego vehicle and simulated agents for safety, correctness of behavior, comfort, and progress.

“Waymax provides a set of intuitive metrics to evaluate the ego vehicle as well as simulated agents for safety and correctness of behavior (such as obeying traffic rules, not colliding), as well as comfort and progress.”

Classified as probably not a substantial claim.

Waymax augments each scenario with feasible paths that the AV could take from its initial position.

“We augment each scenario with feasible paths that the AV could take from its initial position.”

• AssumptionRests on an assumption: The on-route and off-route paths are useful for computing metrics and developing goal-conditioned planning and interactive agents.

Probably not a substantial claim: may not pass S3 (load-bearing).

Waymax runs for a fixed number of steps.

“runs for a fixed number of steps”

Probably not a substantial claim: may not pass S3 (load-bearing).

All metrics in Waymax are computed in closed-loop.

“All metrics in Waymax are computed in closed-loop”

May combine two claims.

The IDM agent in Waymax follows the logged path recorded in the data but uses IDM to adjust the speed profile to avoid collisions and accelerate on free roads.

“The IDM agent in Waymax follows the logged path that is recorded in the data, but uses IDM to adjust the speed profile to avoid collisions and accelerate on free roads.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims. The quote only partly states this claim.

Waymax allows the user to define a dynamics model and provides several pre-defined options for controlling the physical dynamics of vehicles in simulation, including a delta action space and a bicycle action space.

“Waymax allows the user to define a dynamics model and provides several pre-defined options for controlling the physical dynamics of vehicles in simulation”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Waymax is designed as a collection of inter-operable libraries while maintaining fast simulation speed.

“Waymax is designed as a collection of inter-operable libraries while maintaining fast simulation speed.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The main Waymax libraries comprise common data-structures, a distributed data-loading library, simulator components such as metrics and dynamics, and a Gym-like environment interface.

“The main libraries comprise of (1) a set of common data-structures, (2) a distributed data-loading library, (3) simulator components such as metrics and dynamics, and (4) a Gym-like environment interface.”

Classified as probably not a substantial claim.

Each component of the Waymax simulator can be modified, replaced, or used standalone by the user.

“Each component of the simulator can be modified, replaced, or used standalone by the user.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The Waymax state includes static data consisting of road and lane boundaries sampled as a 3D point-cloud roadgraph, as well as on-route and off-route paths for the ego vehicle.

“The static data includes the road and lane boundaries sampled as a 3D point-cloud (known as the “roadgraph”), as well as on-route and off-route paths for the ego vehicle.”

• AssumptionRests on an assumption: The first component of defining autonomous driving as a sequential control problem is defining the state space.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In contrast with stateful simulators such as OpenAI Gym and DM Control, Waymax users need to maintain the simulator state within a simulation loop and interact with the simulator primarily through two functions.

“In contrast with stateful simulators, such as OpenAI Gym [6] and DM Control [32], Waymax users need to maintain the simulator state within a simulation loop and interact with the simulator primarily through two functions:”

• cites [6]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [32]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Waymax runs entirely on hardware accelerators (TPUs/GPUs) and supports in-graph simulation for training.

“It runs entirely on hardware accelerators such as TPUs/GPUs and supports in-graph simulation for training, making it suitable for modern large-scale, distributed machine learning workflows.”

• Also stated in: § 1 Introduction · plain

  “Waymax is written using JAX [5], which allows simulation to be run entirely on accelerators such as graphics and tensor processing units (GPUs and TPUs).”

• Also stated in: § 2 Related Work · plain

  “our simulator is designed to support hardware-accelerated training, the agent models can be connected in-graph for training and inference, and the simulation is differentiable.”

• cites [5]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The paper benchmarks a suite of popular imitation and reinforcement learning algorithms with ablation studies, highlighting the effectiveness of routes as guidance for planning agents and the ability of RL to overfit against simulated agents.

“we benchmark a suite of popular imitation and reinforcement learning algorithms with ablation studies on different design decisions, where we highlight the effectiveness of routes as guidance for planning agents and the ability of RL to overfit against simulated agents.”

May combine two claims.

Waymax is a data-driven, multi-agent simulator for autonomous driving, built on real-world driving data (e.g., WOMD), designed for large-scale simulation and testing.

“we introduce Waymax, a new data-driven simulator for autonomous driving in multi-agent scenes, designed for large-scale simulation and testing.”

• Also stated in: § 1 Introduction · plain

  “we propose Waymax-a differentiable, hardware-accelerated and multiagent simulator that is built using real-world driving data from the Waymo Open Dataset.”

• Also stated in: § 2 Related Work · plain

  “Waymax is a differentiable, hardware-accelerated and multi-agent simulator that is built on top of real-world driving data.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Users can inject multiple agent models and dynamics models into the Waymax environment, where each model can control multiple objects.

“users can inject multiple agent models and dynamics model to Waymax environment where each model can control multiple objects”

Probably not a substantial claim: may not pass S3 (load-bearing).

Users primarily interact with Waymax as a partially-observable stochastic game.

“Users primarily interact with Waymax as a partially-observable stochastic game.”

The quote only partly states this claim.

Waymax is differentiable.

“the simulation is differentiable.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Waymax provides a default observation function that transforms the location of all other vehicles to the agent's own coordinate frame and sub-samples the roadgraph via distance.

“We provide a default observation function that transforms the location of all other vehicles to the agent's own coordinate frame, and sub-samples the roadgraph via distance.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Waymax state includes dynamic data consisting of position, rotation, velocity, and bounding box dimensions for all vehicles, cyclists, and pedestrians in a scene, along with traffic light colors.

“The dynamic data in the state consists of the position, rotation, velocity, and bounding box dimensions for all vehicles, cyclists, and pedestrians in a scene, along with the color of traffic light signals (red, yellow, green).”

• AssumptionRests on an assumption: The first component of defining autonomous driving as a sequential control problem is defining the state space.

May combine two claims.

Waymax enables standard training and evaluation workflows and reliable benchmarking in both open and closed-loop settings.

“Waymax enables standard training and evaluation workflows and reliable benchmarking in both open and closed-loop settings”

Probably not a substantial claim: may not pass S3 (load-bearing).

Waymax supports controlling an arbitrary number of objects in a scene.

“Waymax is a simulator that supports controlling arbitrary number of objects in a scene.”

May combine two claims.

The paper provides implementation of a representative set of IL and RL baselines and reports their performance against a standard set of metrics on Waymax as references.

“we also provide the implementation of a representative set of IL and RL baselines and report their performance against a standard set of metrics on Waymax as references.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Each agent views the simulator state through a user-defined observation function, which can induce partial observability.

“Each agent views the simulator state through a user-defined observation function, which can induce partial observability.”

May combine two claims.

Making predictions at a lower frequency than the environment frequency improved performance, so we predict 5-step long trajectories and only replan every 5 steps.

“We found that making predictions at a lower frequency than the environment frequency improved performance, so we predict 5-step long trajectories and only replan every 5 steps.”

May combine two claims.

When considering batch size 1 and using a GPU, Waymax achieves over 1000Hz for the Step function and over 2000Hz if only considering the Transition.

“When considering batch size 1 and using a GPU, Waymax achieves over 1000Hz for Step function and over 2000Hz if only considering the Transition.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the PlanningAgentEnvironment, all other agents are controlled by user-specified sim agents or log playback.

“All other agents are controlled by user-specified sim agents or log playback (Fig. 3).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Using the reset and step functions, a user can run a simple but complete simulation of a stochastic game between multiple agents.

“Using these two functions, a user can run a simple, but complete simulation of a stochastic game between multiple agents”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Step function computes both the state transition and the reward.

“Note that the Step function computes both the state transition and the reward.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

By construction, the bicycle action space does not violate the comfort metric.

“By construction the bicycle action space does not violate the comfort metric.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Waymax provides adapters to convert the functionally pure Waymax simulator into a stateful one to support existing codebases.

“In addition, we do provide adapters to convert the functionally pure Waymax simulator into a stateful one to support existing codebases.”

May combine two claims.

Running batch size 16 gives an equivalent runtime of over 5000Hz per example, closer to 500 times faster than using a CPU.

“gives an equivalent runtime of over 5000Hz per example (i.e.closer to 500 times faster than using a CPU).”

May combine two claims.

Waymax supports both multi-agent simulation that allows users to control arbitrary objects within the scenario, and a single-agent workflow where a single AV agent is trained using learned or rule-based models to control the other vehicles in the scene.

“Thus, Waymax supports both multi-agent simulation that allows users to control arbitrary objects within the scenario, as well as a single-agent workflow where a single AV agent is trained using learned or rule-based models to control the other vehicles in the scene.”

• AssumptionRests on an assumption: The ultimate goal of the autonomous driving problem is to train an AV planning agent.

May combine two claims.

For open-loop imitation, the discrete action space performs best, possibly because it is easier to model multi-modal behavior.

“For open-loop imitation, the discrete action space performs best, possibly because it is easier to model multi-modal behavior.”

May combine two claims.

The discrete action space outperforms the adapted Wayformer model, likely due to the fact that it is trained explicitly for this task.

“Furthermore, it outperforms the adapted Wayformer model, likely due to the fact that it is trained explicitly for this task.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Rollout function rolls out the environment given an Actor for an entire episode (i.e., 80 steps for WOD).

“We also benchmark a Rollout function which rolls out the environment given an Actor for an entire episode (i.e., 80 steps for WOD).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Following WOMD, the environment controls up to 128 objects in one scene.

“Following WOMD, the environment controls up to 128 objects in one scene.”

• cites WOMDNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The expert using the discrete bicycle action space has comparable performance to the other experts, confirming that the discretization is sufficiently fine.

“The expert using the discrete bicycle action space has comparable performance to the other experts, confirming that the discretization is sufficiently fine.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The MultiAgentEnvironment provides an interface for multi-agent and sim-agent problems.

“The MultiAgentEnvironment provides an interface for multi-agent and sim-agent problems.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The PlanningAgentEnvironment exposes an interface for controlling only the ego vehicle in the scene.

“The PlanningAgentEnvironment exposes an interface for controlling only the ego vehicle in the scene.”

• AssumptionRests on an assumption: The ultimate goal of the autonomous driving problem is to train an AV planning agent.

May combine two claims.

Waymax supports hardware acceleration and distributed training for efficient and cost-effective training of machine-learned models.

“Waymax provides diverse scenarios drawn from real driving data, and supports hardware acceleration and distributed training for efficient and cost-effective training of machine-learned models.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The Metrics function consumes more computation than the Transition function because the Off-Road metric needs to find nearby roadgraph points, which is a slow operation.

“Noticeably the Metrics function consumes more computation then the Transition function because the Off-Road metric needs to find nearby roadgraph points, which is a slow operation.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Waymax is written as a collection of inter-operable libraries for data loading, metric computation, and simulation, which can support a wide variety of research problems not limited to planning evaluations.

“It is also designed with flexibility in mind -Waymax is written as a collection of inter-operable libraries for data loading, metric computation, and simulation, which can support a wide variety of research problems that are not limited to just the planning evaluations presented in this work.”

Probably not a substantial claim: may not pass S3 (load-bearing).

We adapt the state-of-the-art Wayformer behavior prediction model to the planning setting.

“As a point of reference, we adapt the state-of-the-art Wayformer behavior prediction model [35] to the planning setting.”

• cites [35]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Waymax supports hardware acceleration on GPUs and TPUs, and supports in-graph training where training and simulation run entirely on the accelerator without host communication bottlenecks.

“Waymax supports both hardware acceleration on GPUs and TPUs, as well as combining training and simulation within the same computation graph (referred to as”

Probably not a substantial claim: may not pass S3 (load-bearing).

We provide a number of expert agent models to provide groundtruth actions for open-loop training.

“Expert We provide a number of expert agent models to provide groundtruth actions for open-loop training.”

May combine two claims.

The stateless design of Waymax enables efficient optimization through JAX's JIT compiler and functional libraries, and easily allows users to implement control algorithms that require backtracking, such as search.

“This stateless design enables efficient optimization through JAX's [5] JIT compiler and functional libraries, and easily allows users to implement control algorithms that require backtracking, such as search.”

• cites [5]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The Waymo Open Motion Dataset includes over 100,000 trajectory snippets and 7.64 million unique objects to interact or control.

“which includes over 100, 000 trajectories snippets and 7.64 million unique objects to interact or control.”

• cites [15]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Together, the feasible paths describe all the ways in which the AV can legally drive in the scenario.

“Together, these paths describe all the ways in which the AV can legally drive in the scenario.”

Classified as probably not a substantial claim. May combine two claims.

The driving policy modeling problem is primarily focused on behavior than perception, and therefore the paper does not intend to support sensor simulation.

“It is worth noting that the driving policy modeling problem is primarily focused on behavior than perception, thereby, we do not intend to support sensor simulation, such as [14,28,48,9,2], which represents another important research area for autonomous driving perception.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Waymax data loading and processing can be extended to other popular datasets without loss of generality.

“Waymax data loading and processing can be extended to other popular datasets without loss of generality.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Behavior cloning suffers from covariate shift and causal confusion.

“It is widely known that BC suffers from covariate shift [42] and causal confusion [13].”

• cites [42]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [13]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The route progress ratio can be greater than 1.

“Since the vehicle can continue driving after reaching its destination, this ratio could be greater than 1.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Open-loop imitation learning (behavior cloning) has been applied to predict driving behaviors of other road users and the ego vehicle.

“open-loop imitation learning (IL), known as behavior cloning (BC), has been applied to predict driving behaviors of other road users [3,10,49,26,36,17,34] as well as the ego vehicle [39,4,41,12,53,11].”

• cites [3]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [10]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [49]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [26]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [36]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [17]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [34]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [39]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [4]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [41]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

A major simulation challenge remains in the generation of diverse scenarios and realistic behavior for other agents, and behavior challenges have been shown to be a significant bottleneck to scaling.

“However, a major simulation challenge still remains in the generation of diverse scenarios and realistic behavior for other agents (such as vehicles and pedestrians) in the scene, and as the driving field has matured, behavior challenges have been shown to be a significant bottleneck to scaling [31].”

• cites [31]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Each WOMD trajectory snippet is 9 seconds recorded with 0.1 Hz.

“Each trajectory snippet is 9 seconds recorded with 0.1 Hz.”

• cites [15]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Users who only need one component of Waymax, or who wish to significantly modify the simulator's behavior, can easily do so through Waymax's APIs.

“users who only need one component of Waymax (e.g.only metrics, or only data loading), or who wish to significantly modify the behavior of the simulator (such as generating synthetic scenerios) can easily do so through Waymax's APIs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Simulators such as CARLA generate synthetic scenarios.

“simulators that generate synthetic scenarios (e.g.CARLA [14])”

• cites [14]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Closed-loop methods including adversarial imitation learning and reinforcement learning have been proposed to address these challenges by learning from feedback and explicit hand designed rewards in the simulator.

“Closedloop methods including adversarial imitation learning [20,7] and reinforcement learning (RL) methods [22,21,54,27] have been proposed to address these challenges by learning from feedback and explicit hand designed rewards in the simulator.”

• cites [20]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [7]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [22]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [21]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [54]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [27]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Originally, the Wayformer predicts multiple 8-second future trajectories given a 1-second context history.

“Originally, the Wayformer predicts multiple 8-second future trajectories given a 1-second context history.”

• cites [35]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The nominal values represent noise in the bounding boxes and logged data and serve as a lower bound for performance.

“The nominal values represent noise in the bounding boxes and logged data and serve as a lower bound for performance.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

OpenAI Gym and DM Control are stateful simulators.

“stateful simulators, such as OpenAI Gym [6] and DM Control [32]”

• cites [6]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [32]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

IDM describes a rule for updating the acceleration of a vehicle to avoid collisions based on the proximity and relative velocity of the vehicle to the object directly in front of the vehicle.

“IDM describes a rule for updating the acceleration of a vehicle to avoid collisions based on the proximity and relative velocity of the vehicle to the object directly in front of the vehicle”

• cites [51]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

Putting multiple policies in one environment directly is not a flexible way because it is hard to coordinate different policies or change policies.

“While it might be possible to put multiple policies in one environment directly, it is certainly not a flexible way as it is hard to coordinate different policies or change policies.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Hyperparameter selection for all learned benchmark agents (BC, DQN) was performed via a grid search.

“We performed hyperparameter selection for all learned benchmark agents (BC, DQN) outlined in Table 3 via a grid search.”

Classified as probably not a substantial claim.

JAX converts operations into XLA, a linear algebra instruction set and optimizing compiler that supports execution on CPU, GPU, or TPU.

“which converts operations into XLA [43], a linear algebra instruction set and optimizing compiler which supports execution on CPU, GPU, or TPU.”

• cites [43]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The RL agent tends to overfit or exploit the behavior of 'easier' IDM agents.

“We believe this is because the RL agent tends to overfit or exploit the behavior of 'easier' IDM agents.”

• AssumptionRests on an assumption: IDM will stop for the SDC to avoid collisions.

Probably not a substantial claim: may not pass S3 (load-bearing).

This serves as a check that the Waymax environment is producing the correct training data.

“This serves as a check that the Waymax environment is producing the correct training data.”

Classified as probably not a substantial claim.

The sim-to-real gap remains an important limitation for any simulation-based framework.

“While in Waymax we have made design decisions to minimize this gap (such as using real-world data to seed scenarios), this remains an important limitation for any simulation-based framework.”

• AssumptionRests on an assumption: The problem of sim-to-real transfer is a critical issue in autonomous driving, and it is difficult to guarantee that the same performance and level of safety will carry over to the real world.