PaperlayerView exact revision on arXiv

Source-linked AI summary

CaRL: Learning Scalable Planning Policies with Simple Rewards

Bernhard Jaeger, Daniel Dauner, Jens Beißwenger, Simon Gerstenecker, Kashyap Chitta, Andreas Geiger

arXiv:2504.17838v3cs.LGcs.AIcs.RO

TL;DR

Privileged-driving RL faces reward-design challenges: common multi-term rewards can impede PPO scalability. CaRL trains with route completion and infraction penalties, achieving strong results across CARLA and nuPlan benchmarks.

  • Problem

    Rule-based privileged planners require scenario-specific rules for complex driving, while multi-term RL rewards introduce tradeoffs, local minima, and undesirable shortcuts.

    Source: §1 · p. 1

  • Method

    CaRL uses route completion as its reward source, ending episodes for major infractions and multiplicatively reducing route completion during soft-constraint violations.

    Source: §1 · p. 2

  • Results

    Across CARLA and nuPlan, CaRL outperforms prior RL methods and is the best learning-based nuPlan approach; on longest6 v2 it beats Roach by 42 DS.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The route-completion reward lets PPO use larger minibatches, supporting efficient scaling to more training samples through parallel data collection.

    Source: §5 · p. 8

    Sim2Real transfer to real cars remains to be demonstrated and is left for future work.

    Source: §5 · p. 9

Abstract

from arXiv · show

We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL approaches for driving use complex shaped rewards that sum multiple individual rewards, \eg~progress, position, or orientation rewards. We show that PPO fails to optimize a popular version of these rewards when the mini-batch size is increased, which limits the scalability of these approaches. Instead, we propose a new reward design based primarily on optimizing a single intuitive reward term: route completion. Infractions are penalized by terminating the episode or multiplicatively reducing route completion. We find that PPO scales well with higher mini-batch sizes when trained with our simple reward, even improving performance. Training with large mini-batch sizes enables efficient scaling via distributed data parallelism. We scale PPO to 300M samples in CARLA and 500M samples in nuPlan with a single 8-GPU node. The resulting model achieves 64 DS on the CARLA longest6 v2 benchmark, outperforming other RL methods with more complex rewards by a large margin. Requiring only minimal adaptations from its use in CARLA, the same method is the best learning-based approach on nuPlan. It scores 91.3 in non-reactive and 90.6 in reactive traffic on the Val14 benchmark while being an order of magnitude faster than prior work.

1 Introduction

Privileged-driving RL is promising, but complex reward tradeoffs limit PPO scalability; CaRL instead optimizes route completion and penalizes infractions. It scales to large minibatches and outperforms prior RL planners on CARLA and learning-based baselines on nuPlan.

1 Introduction

  • Complex driving rewards combine terms whose tradeoffs require tuning, introduce local minima, and can encourage undesirable shortcuts.The authors report that increasing minibatch size fourfold drastically reduces PPO performance with a complex reward, consistent with optimization getting stuck in a local optimum.Source: §1 · p. 2
  • CaRL uses route completion as its reward source, ending episodes for major infractions and multiplicatively reducing route completion during soft-constraint violations.The design does not rely on rule-based planners to compute reward terms.Source: §1 · p. 2
  • Increasing minibatch size from 256 to 1024 improves learning with the simple reward, which outperforms the complex Roach reward.The complex reward's performance instead drops drastically as minibatch size increases.Source: §1 · p. 2, Figure 1 · p. 2
  • With minibatch size 16384, CaRL trains on 300 million CARLA samples—30x prior work—using one compute server.The authors attribute the ability to train on more data to large minibatches that enable efficient parallel data collection.Source: §1 · p. 2
  • On longest6 v2, CaRL exceeds Roach by 42 DS and Think2Drive by 57 DS; on nuPlan Val14 it scores 91 closed-loop in both traffic modes.It also outperforms Diffusion Planner by 1.7 points in non-reactive traffic and 7.9 points in reactive traffic, with 10x faster inference.Source: §1 · p. 2

2 Improving Scalability for RL

This section establishes a CARLA training and evaluation setup, then adapts PPO hyperparameters to support faster, stronger training. The Atari settings reduce off-policy optimization steps and training time, while overly high learning rates can produce degenerate policies.

2.1 Preliminaries

  • Longest6 v2 uses 36 routes in towns 1–6, with 5–21 pre-crash safety-critical scenarios per route; its faster towns support quicker RL experimentation.The evaluation averages five training seeds, each evaluated three times, unless noted otherwise.Source: §2.1 · p. 3

2 Improving Scalability for RL

  • The training simulator avoids town reloads and preprocesses A* route planning, enabling Roach reproduction with 10 million samples in 32 hours on one A100 GPU.Evaluation uses the original leaderboard 2.0 code for fair comparisons.Source: §2.1 · p. 3

2.2 Enabling higher learning rates

  • PPO's Atari hyperparameters provide a basis for higher learning rates, reducing training time and improving Roach performance relative to its standard CARLA settings.PPO performs multiple gradient steps per data collection iteration, approximating an off-policy policy gradient; clipping and limiting update steps control the approximation error.Source: §2.2 · pp. 3–4
  • With the Atari learning rate and schedule, Roach's DS falls from 22 ± 14 to 2 ± 1 when other hyperparameters remain unchanged.Roach's 959 off-policy steps are 60 times the Atari count, but its learning rate is 25x lower, balancing the resulting policy changes.Source: §2.2 · p. 4
  • The Atari hyperparameters train 10 hours faster and achieve 11 DS better performance than Roach's CARLA settings.The time reduction comes from using five times fewer epochs; the performance gain may reflect a combination of hyperparameters.Source: Table 2 · p. 4, §2.2 · p. 4
  • Using the action-distribution mean at inference achieves 1 DS better performance and reduces standard deviation by 3 points versus using its mode.The authors attribute the mode's bias to PPO optimizing expected performance under the action distribution, which is its mean.Source: §2.2 · p. 4

3 Optimizing a Single Reward

CaRL replaces dense, rule-derived rewards with route completion and infraction penalties, and this simpler objective lets PPO improve at larger mini-batch sizes and scale to more training data.

3 Optimizing a Single Reward

  • Dense rewards based on handcrafted desired speed, position, and orientation offer rich feedback but can introduce local minima, exploitable shortcuts, and suboptimal performance.With the Roach reward, a policy sometimes waits for a red light rather than driving through a green one.Source: §3 · pp. 4–5
  • CaRL rewards route completion multiplied by soft-penalty factors, while major infractions such as collisions end the episode.Soft penalties can be violated to avoid hard penalties; comfort penalties must remain above zero when unavoidable early in training.Source: §3 · p. 5
  • The reward has finite total return, avoids prescribing a driving method, and shares its global optimum with the evaluation metric.The agent receives optimal reward by completing the route without infractions, which also yields optimal DS.Source: §3 · p. 5
  • At batch size 1024, PPO with the simple reward improves while PPO with the dense Roach reward can get stuck; Table 3 reports DS.At batch size 256, the simple reward drops 13 points; increasing its batch size to 1024 improves DS by 17 points, beats the Roach reward by 4 DS, and reduces variance from ±7 to ±3.Source: §3 · pp. 5–6, Table 3 · p. 6

3.1 Scaling Data

  • Scaling PPO to 300 million samples with a larger mini-batch improved performance by 33 DS, using one 8-GPU node.The 300M result averages three seeds; prior driving PPO work used 1–10 million samples.Source: §3.1 · p. 6
  • Large mini-batches allow data collection to be parallelized, supporting the move from 10 million to 300 million samples on one 8-GPU node.The scaling experiment used DD-PPO with parallel simulators and ran for one week.Source: §3.1 · p. 6

4 Experiments

Experiments compare CaRL with planning baselines on CARLA longest6 v2 and nuPlan Val14. CaRL leads the learning-based methods on both benchmarks, while rule-based planners retain the strongest scores.

4 Experiments

  • The experiments compare methods under the same settings; CARLA baselines include Roach, Think2Drive, PDM-Lite, and PlanT.The planners are privileged, using ground-truth perception inputs.Source: §4 · p. 6, §4.1 · pp. 6–7

4.1 CARLA

  • 64 DS on longest6 v2; CaRL beats Roach by 42 DS, Think2Drive by 57 DS, and PlanT by 2 DS, while PDM-Lite leads overall at 73 DS.CaRL also particularly reduces pedestrian and vehicle collisions compared with other RL baselines.Source: Table 5 · p. 7, §4.1 · p. 7
  • CaRL runs 2 times faster than PDM-Lite on CARLA, with inference compute similar to other RL methods despite substantially better driving performance.Runtime includes preprocessing observations and was measured on the last longest6 v2 route.Source: §4.1 · p. 7

4.2 nuPlan

  • 91.3 CLS non-reactive and 90.6 CLS reactive; CaRL leads learned planners on Val14 and runs 7–17× faster than the baselines.PDM-Closed exceeds 92 reactive and non-reactive CLS.Source: Table 6 · p. 8, §4.2 · p. 8
  • Because nuPlan runs for a fixed 15 seconds after route completion, training adds a per-frame survival bonus to encourage avoiding infractions after reaching the goal.CaRL is trained for 500M samples on nuPlan.Source: §4.2 · p. 8

5 Conclusion

CaRL uses route completion as its primary reward and scales PPO training to large sample counts, with improved benchmark performance. The paper’s conclusions are bounded by its simulation settings and several untested or identified limitations.

5 Conclusion

  • Using a mini-batch size of 16,384, PPO scales to 300M CARLA samples, improving longest6 v2 performance by 33 DS; CaRL achieves 91 CLS on nuPlan Val14 in non-reactive and reactive traffic.Source: §5 · p. 8
  • CaRL rewards route completion; major infractions terminate episodes, while soft-constraint violations multiplicatively reduce route completion.Examples include collision as a major infraction and exceeding the speed limit as a soft constraint violation.Source: §5 · p. 8
  • The study covers urban driving up to 80 km/h, not highway-specific high-speed driving; transfer from simulation to real cars also remains undemonstrated.The authors leave Sim2Real transfer for future work.Source: §5 · p. 9
  • CaRL trains on seven longest6 v2 scenario types, not all CARLA leaderboard 2.0 scenarios; level 5 generalization to new towns is untested.The evaluated benchmarks are level 4, where training on the evaluation town is allowed.Source: §5 · p. 9
  • Wide comfort bounds allow compliance with the comfort infraction, but the resulting behavior might not be comfortable in a real car.The authors note the absence of human reference values for comfortable driving in CARLA.Source: §5 · p. 9
  • Reported CARLA failure modes are missed highway off-ramps and rear-end collisions attributed to CaRL when another driver runs a red light.The paper notes that CARLA counts rear-end collisions as the agent’s fault.Source: §5 · p. 9
  • Route completion does not reward faster arrival; PPO’s Monte-Carlo returns may help with sparse local feedback, but Q-learning alternatives remain untested.The authors identify algorithms relying on local TD predictions as a possible harder case for optimizing this reward.Source: §5 · p. 9

A Related work

Prior driving RL often adds dense reward signals or expert-derived targets to simplify learning, but these designs can constrain performance. The paper positions CaRL’s single-reward approach alongside work on simulators, scaling, and asynchronous collection.

A Related work

  • Simple speed-and-termination rewards were used in actor-free settings; in reactive traffic, a reward without progress can favor standing still unless behavior cloning or log-replay behavior changes the outcome.The paper describes Lu et al.’s reward as lacking a progress term and notes the additional training or traffic conditions under which driving emerged.Source: A Related work · pp. 16–17
  • Dense rewards simplify credit assignment, but expert-derived desired states can cap performance; CaRL instead removes reward tradeoffs by optimizing a single component.The paper describes desired-state rewards as the dominant paradigm for CARLA RL planners.Source: A Related work · pp. 17–18
  • Roach’s open-source training code was incompatible with the safety-critical scenario runner, so the authors rewrote training and reproduced it on longest6 v2.They report Roach scoring below rule-based and imitation-learning baselines, then show PPO can optimize their reward with a larger mini-batch.Source: A Related work · p. 17
  • Think2Drive uses world-model data to improve sample throughput, but the paper argues CARLA throughput can rise two orders of magnitude through engineering, hyperparameters, and scale.The authors characterize world-model use as a tradeoff between sample efficiency and lower asymptotic policy performance.Source: A Related work · pp. 17–18
  • GIGAFLOW’s nuPlan reactive score is 93.8 versus CaRL’s 93.1, while CaRL uses 1B rather than 1T samples; both use the same learning algorithm.The paper notes that GIGAFLOW’s faster simulation yields similar overall compute requirements despite the sample-count difference.Source: A Related work · p. 18
  • The proposed asynchronous collection avoids Envpool’s bias toward faster, easier environments; AC-PPO synchronizes backward training to avoid policy lag.The text also notes that fully asynchronous methods leave the effect of policy lag on fixed-compute comparisons unclear.Source: A Related work · p. 19

B Method changes

  • Increasing field of view helps the final model exceed 50 DS on longest6 v2, but hurts performance in low-data regimes by making the task harder.The larger view yields smaller objects and more pixels; changes had little impact for the 1024 mini-batch model already near its limited-FOV performance.Source: B Method changes · p. 19

B.1 Asynchronous data collection with PPO

The section introduces asynchronous collection to avoid PPO’s synchronization inefficiency in variable-duration environments, then reports training changes for CARLA. AC-PPO improves collection speed, while reducing training epochs also improves driving score.

B.1 Asynchronous data collection with PPO

  • AC-PPO removes collection synchronization by running forward passes concurrently at batch size 1, avoiding idle time across variable-duration environments.PPO’s batched forward passes synchronize environments after each step, so all wait for the slowest environment.Source: §B.1 · p. 19
  • AC-PPO trains 11% faster than PPO on the same hardware while performing the same computations (up to random seeds), reducing training time by 3 hours.The training code was re-implemented in C++ using the LibTorch interface of PyTorch.Source: §B.1 · p. 20

B.2 Hyperparameter

B.3 Architecture

The architecture builds on Roach’s convolutional and MLP design, with layer normalization added before each ReLU. The action distribution uses a constrained Beta parameterization, and the section reports architecture and distribution ablations.

B.3 Architecture

  • The network encodes BEV semantic images with 6 convolutional layers, processes measurements and merged features with 2-layer MLPs, and predicts Beta parameters and value.Layer normalization is added before every ReLU activation.Source: §B.3 · pp. 20–21
  • 2 points in driving score and 9 in route completion: layer normalization improved both measures.The authors hypothesize that layer normalization may increase network plasticity, which matters in non-stationary on-policy RL.Source: §B.3 · p. 21, Table 10 · p. 21
  • 1 point lower driving score and 1 point higher route completion: the asymmetric critic receives extra value-function inputs unavailable to the policy.Those inputs include route length and time remaining until selected penalties or timeouts.Source: §B.3 · p. 21

B.4 Action distribution

  • Constraining Beta parameters to values of at least 1 keeps the action distribution unimodal and avoids its degeneration toward extreme actions.The Beta distribution is used for bounded continuous actions; Normal distributions performed worse in prior work and preliminary experiments.Source: §B.4 · p. 21
  • The driving-score comparison does not measure comfort, so the two action-distribution variants’ equal scores cannot establish a comfort difference.Figure 2 illustrates example Beta distribution shapes.Source: §B.4 · p. 21, Figure 2 · p. 22

B.5 Input

The input design uses a BEV semantic image and route conditioning, while adjusting field of view and resolution. For nuPlan, the method adapts the inputs and integrates vehicle control into the planner.

B.5 Input

  • The BEV input encodes road, route, markings, traffic controls, and other road users; vehicle and pedestrian bounding boxes also encode speed.The design omits historical actor-state channels aside from vehicle speed and renders selected vehicle signals and a constant-velocity forecast.Source: §B.5 · p. 22, Figure 3 · p. 23
  • Conditioning on the A⋆ route only inside intersections reduced preset lane-change collisions, but sometimes led the agent to take the wrong turn.With route conditioning elsewhere, the policy often followed the planner’s lane choice and could lane-change at predefined locations.Source: §B.5 · p. 22
  • The CARLA observation expands visible range to 78 m in front, 50 m behind, and 64 m to either side, while reducing resolution to 2 pixels per meter.The network adds a strided convolutional layer to keep the CNN output feature count constant.Source: §B.5 · p. 22
  • 10 points lower driving score: the larger-field-of-view, intersection-only-conditioned input reduced vehicle collisions by 0.5 but reduced route completion by 31 and increased route deviations by 0.3.The authors attribute the collision reduction to seeing vehicles earlier and choosing when to lane-change; they say scaling samples and minibatch size may recover lost route completion.Source: Table 12 · p. 23, §B.5 · p. 23

B.6 nuPlan

  • nuPlan training used 500M/1B samples, enabled by the simulator’s lower memory and compute footprint and fewer training crashes.The architecture is the same as in CARLA apart from input-layer dimensions.Source: §B.6 · p. 23
  • The nuPlan planner learns control directly rather than predicting a trajectory for a separate controller to follow.Its action distribution maps to target acceleration and steering, which are converted to bicycle-model inputs using a discrete-time derivative for steering rate.Source: §B.6 · p. 24

C.1 Roach

The appendix describes Roach reward and route-planning failure cases, CaRL’s CARLA and nuPlan failures, and engineering constraints on simulator throughput. It reports that CARLA server crashes remain a major bottleneck at scale.

C.1 Roach

  • A⋆-prescribed lane changes can force Roach to stop for oncoming traffic or collide during a lane change, creating rear-end risk on multi-lane roads.The figure illustrates predefined lane changes leading to rear collisions.Source: §C.1 · p. 24, Figure 4 · p. 25
  • Waiting at a green light can earn Roach high rewards because it may collect speed rewards after the light turns red; some later seeds still slow near green lights.An early model displayed the behavior after roughly 2 million samples.Source: §C.1 · pp. 24–25, Figure 5 · p. 25

C.2.1 CARLA

  • CaRL’s common CARLA failures include being hit by a fast red-light runner from behind and missing turns, especially highway off-ramps.In the collision example, CaRL tries to evade into the opposing lane but accelerates too slowly.Source: §C.2.1 · pp. 25–26

C.2.2 nuPlan

  • In nuPlan, collision avoidance remains difficult with non-reactive or abruptly behaving reactive vehicles, and pedestrians challenge the planner’s perception and flexible-motion handling.Reactive traffic also makes route completion harder when IDM vehicles block the road.Source: §C.2.2 · p. 26

D Engineering; D.1 CARLA

  • Simulator engineering enabled scaling to 300 million CARLA samples and 1 billion nuPlan samples, but frequent CARLA server crashes roughly doubled training time.The authors identify fixing simulator bugs and optimizing slow HD-map retrieval as promising future work.Source: D Engineering · p. 26, §D.1 · p. 27

D.2 Scenario Generation

The appendix describes how the authors generated CARLA training routes and adapted their data, hardware, and PlanT baseline for evaluation. The generated scenarios were imperfect but supplied the volume of unique routes needed for PPO training at scale.

D.2 Scenario Generation

  • The CARLA script builds 1 km routes from HD-map points, validates them, and attempts to place random scenarios roughly every 100 meters.Generated routes contain up to 10 scenarios; imperfect rejection tests leave some routes noisy and scenario types unbalanced, but the script supplies unique training data at scale.Source: §D.2 · pp. 27–28

D.3 nuPlan

  • On nuPlan, scenario generation was unnecessary, while the dataset pipeline compressed selected scenarios and capped nearby entities to reduce database and rendering costs.The dataset uses 269,075 temporally non-overlapping 15-second scenarios; reactive traffic and observation rendering remain performance bottlenecks.Source: §D.3 · p. 28, §D.2 · p. 28

D.4 Hardware

  • The experiments use servers with 12–24 physical CPU cores per GPU because these highly parallelized jobs are CPU-bottlenecked.The authors caution that machines with as few as four CPU cores per GPU run several times slower; their A100 and H100 GPUs were not fully utilized.Source: §D.4 · p. 28

E.1 CARLA

  • The appendix reproduces PlanT and Think2Drive as CARLA baselines.The following PlanT adaptation keeps its architecture largely unchanged while adding input classes and updating control for faster driving.Source: §E.1 · p. 29, §E.1.1 · p. 29

E.1.1 PlanT

  • PlanT inputs add pedestrians, static obstacles, and trigger boxes for relevant stop signs and traffic lights; its model adds four class embeddings and doubles speed bins.The control logic is also replaced with PDM-Lite lateral control; increasing waypoint count to eight or sampling at 2 Hz performed poorly.Source: §E.1.1 · p. 29
  • PlanT is trained on about 470,000 samples covering all 36 leaderboard 2.0 scenarios, for 47 epochs with batch size 128 and learning rate 1e-4.The dataset is collected using PDM-Lite and is similar in size to the 2x dataset used in the original paper.Source: §E.1.1 · p. 29

E.1.2 Think2Drive

The Think2Drive reproduction combines a DreamerV3 world model with actor-critic learning and a weighted reward. Its training uses staged routes, parameter resets, and replay-buffer sampling that emphasizes frames near episode termination.

E.1.2 Think2Drive

  • Think2Drive combines the DreamerV3 model-based reinforcement learning algorithm with a world model and actor-critic approach.The authors reimplemented it because its code was not officially released and inferred several unspecified hyperparameters.Source: §E.1.2 · p. 29
  • The reproduction uses 30 discrete actions at 10 Hz, repeated twice in a 20 Hz simulator, and samples four historical BEV frames as input.The input frames are current and three earlier timesteps; the model has 100 million parameters to match Think2Drive’s model capacity.Source: §E.1.2 · pp. 29–30
  • Its reward is a weighted sum of speed, travel, route-deviation, and steering terms, with weights 1, 1, 2, and 0.5, respectively.The target speed is the minimum of limits based on speed limits, traffic lights, stop signs, and vehicles ahead on the route.Source: §E.1.2 · p. 30
  • Episodes terminate for red lights, stop signs, collisions, or route deviation over 15 m, with terminal rewards ranging from +1 at route completion to negative penalties for infractions.A collision or traffic-sign infraction receives −1 − speed; exceeding the route-deviation threshold receives −1.Source: §E.1.2 · p. 30
  • Training proceeds from 400,000 basic-route steps to 1.5 million steps on 1 km scenario routes, with actor and critic reinitialized at phase transition and periodically afterward.The actor-critic train ratio rises from 16 to 64; half of sampled replay frames come from the 64 frames immediately preceding episode termination.Source: §E.1.2 · p. 31

E.2 nuPlan

The nuPlan section compares planners and describes how the reward is adapted to nuPlan’s scoring and short scenarios. A tuned reward improves CaRL’s scores enough to outperform the rule-based PDM-Closed baseline in both traffic modes.

E.2 nuPlan

  • CaRL closes the reactive-traffic gap to PDM-Closed to 1.5 points, while rule-based and hybrid planners outperform imitation-learning baselines.Diffusion Planner, the best open-source method in that comparison, remains 10 CLS points below PDM-Closed in reactive traffic.Source: §E.2 · p. 32
  • Human-log performance is 93.5 with nuPlan’s standard LQR controller and 96.24 with iLQR, which improves collisions and drivable-area compliance.The iLQR controller raises the reference score by almost 3 CLS points; perfect tracking adds only 0.15 over iLQR.Source: §E.2 · p. 32
  • 93.87 CLS in non-reactive and 93.12 CLS in reactive traffic let tuned CaRL outperform PDM-Closed in both modes.The tuned reward adds 2.6 points in non-reactive traffic and 2.5 points in reactive traffic over CaRL, aligning its terms with nuPlan’s score tradeoffs.Source: §E.2 · p. 32, Table 14 · p. 33, Table 15 · p. 33

F Reward

  • CaRL’s reward centers on route completion: six infractions terminate episodes, while soft violations multiplicatively reduce route completion.The terminal penalty is 1.0 for red-light violations and collisions and 0.0 for the other infractions.Source: F Reward · p. 33, §F.1 · pp. 33–34, §F.2 · p. 34

F.2 Soft Penalties

  • Comfort thresholds are subjective and were set loosely in CARLA; comfort and TTC penalties persist for 500 simulator steps there, but until episode end in nuPlan.The authors leave tuning comfort bounds for more comfortable policies to future work.Source: §F.2 · p. 34

F.3 nuPlan Specific Adaptations

  • For nuPlan, red-light, stop-sign, route-deviation, and blocked terminal penalties are omitted because of label noise, metric mismatch, or the 15-second simulation duration.The adaptation retains off-road termination and modifies lane-center and outside-lane soft penalties; it uses stricter comfort boundaries based on human trajectories.Source: §F.3 · p. 35
  • When route completion reaches 100% before a short nuPlan simulation ends, CaRL adds a constant survival bonus because route-completion reward no longer encourages collision avoidance.The reward blend uses a survival ratio of 0.6 and keeps the sum of rewards in the 0–100 range.Source: §F.3 · p. 35
  • NuPlan failure examples include pedestrian collisions and veering off-road, rear-side collisions during unprotected turns, and reactive traffic blocking progress.The examples distinguish failures in non-reactive scenarios from blocked-road cases in reactive simulation.Source: Figure 9 · p. 37
Loading 2504.17838v3…