PaperlayerView exact revision on arXiv

Source-linked AI summary

BITS: Bi-level Imitation for Traffic Simulation

Danfei Xu, Yuxiao Chen, Boris Ivanovic, Marco Pavone

arXiv:2208.12403v1cs.ROcs.LG

TL;DR

Realistic road-user behavior remains a gap in autonomous-vehicle simulation, making validation reliant on costly real-world testing. BITS learns hierarchical traffic behavior from driving logs and adds planning for stable long-horizon rollouts. Across two real-world AV datasets, it generates diverse, realistic, and stable simulations.

  • Problem

    Realistic road-user behavior remains difficult to simulate, so AV developers still rely on large-scale real-world testing for validation and verification.

    Source: §1 · p. 1

  • Method

    BITS learns spatial goal distributions and goal-conditioned control from driving logs, then uses prediction and planning to stabilize long-horizon behavior.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    On two large-scale real-world AV datasets, BITS generates diverse, realistic, and stable traffic simulations.

    Source: §5 · p. 10

  • Takeaways & Limitations

    The authors provide evaluation metrics and an open-source tool for unifying dataset formats and converting scenes into interactive simulations as a foundation for further research.

    Source: §1 · p. 2

    Data-driven simulation models inherit geographic biases from their training data, potentially raising safety concerns in underrepresented regions.

    Source: §5 · pp. 9–10

Abstract

from arXiv · show

Simulation is the key to scaling up validation and verification for robotic systems such as autonomous vehicles. Despite advances in high-fidelity physics and sensor simulation, a critical gap remains in simulating realistic behaviors of road users. This is because, unlike simulating physics and graphics, devising first principle models for human-like behaviors is generally infeasible. In this work, we take a data-driven approach and propose a method that can learn to generate traffic behaviors from real-world driving logs. The method achieves high sample efficiency and behavior diversity by exploiting the bi-level hierarchy of driving behaviors by decoupling the traffic simulation problem into high-level intent inference and low-level driving behavior imitation. The method also incorporates a planning module to obtain stable long-horizon behaviors. We empirically validate our method, named Bi-level Imitation for Traffic Simulation (BITS), with scenarios from two large-scale driving datasets and show that BITS achieves balanced traffic simulation performance in realism, diversity, and long-horizon stability. We also explore ways to evaluate behavior realism and introduce a suite of evaluation metrics for traffic simulation. Finally, as part of our core contributions, we develop and open source a software tool that unifies data formats across different driving datasets and converts scenes from existing datasets into interactive simulation environments. For additional information and videos, see https://sites.google.com/view/nvr-bits2022/home

1 Introduction

Realistic road-user behavior remains a major gap in autonomous-vehicle simulation, leaving developers reliant on costly and potentially dangerous road testing. BITS learns traffic behavior from driving logs using hierarchical intent inference, goal-conditioned control, and planning, and is evaluated on two driving datasets.

1 Introduction

  • Realistic road-user behavior remains a critical simulation gap, so AV developers still rely on large-scale real-world testing for validation and verification.Testing directly on roads is costly and potentially dangerous.Source: §1 · p. 1
  • Log replay cannot react to counterfactual ego motions, while heuristic controllers struggle to generate diverse, complex behaviors such as yielding and cutting-in.These limitations make it difficult to validate new features across real-world scenarios.Source: §1 · pp. 1–2
  • Learning from logs grounds reactive behavior in real driving, but prediction errors can compound in long-horizon closed-loop simulation, causing divergence, collisions, or road departures.Learned models may also be brittle under domain shifts and face combinatorial multi-agent state spaces.Source: §1 · p. 2
  • BITS separates spatial intent inference from goal-conditioned control, then uses prediction and rule-based planning to balance human-like in-distribution behavior with reduced out-of-distribution divergence.The authors evaluate BITS on Lyft Level 5 and nuScenes, introduce realism metrics, and release a dataset-format and scene-conversion tool.Source: §1 · p. 2

2 Related work

Prior traffic simulation spans analytical controllers, learned simulation models, trajectory prediction, and hierarchical imitation. BITS positions its agent-centric, goal-conditioned approach as a way to generate diverse, stable, long-horizon closed-loop behavior.

2 Related work

  • Analytical microscopic simulators use fixed routes and simplified motion models, limiting the complexity and expressiveness needed to develop or evaluate autonomous-driving features.The paper focuses on microscopic simulation because it concerns individual agents and their interactions.Source: §2 · p. 2
  • Unlike STRIVE’s worst-case, short open-loop scenarios and TrafficSim’s coordinated scene-level control, BITS targets broad long-horizon closed-loop behaviors with independent agent-level decisions.The framework overview pairs goal-conditioned action candidates with neighbor-trajectory forecasts and rule-based cost selection.Source: §2 · p. 3, Figure 1 · p. 3
  • Trajectory prediction models perform well over short horizons but generally degrade over longer ones; goal-conditioned multi-stage methods first infer goals and then connect current positions to them.BITS applies goal inference to closed-loop simulation rather than only open-loop prediction.Source: §2 · p. 3
  • Hierarchical policies combine high-level goals with low-level execution, while multimodal imitation can represent alternative behaviors; BITS uses 2D bird’s-eye-view maps to represent driving goals [55] [58].The paper reports realistic, diverse, stable long-horizon simulations.Source: §2 · p. 3

3 Bi-level imitation for traffic simulation

BITS treats traffic simulation as agent-centric imitation learning and uses a bi-level policy to generate diverse behaviors. Prediction-and-planning supports stable long-horizon rollouts, while analytical and learned metrics address simulation evaluation without a single ground truth.

3 Bi-level imitation for traffic simulation

  • The section introduces a hierarchical imitation framework, a prediction-and-planning module for long-horizon stability, and analytical and learned simulation metrics.Implementation details are deferred to the experiments and appendix.Source: §3 · p. 4

3.1 Traffic simulation as imitation learning

  • Each agent makes decentralized decisions without explicit coordination, while real-world logs provide demonstrations; partial observability makes actions ambiguous and independent agents create a combinatorial future-state space.The model uses each agent’s state and decision context, including local map information and histories of neighboring agents.Source: §3.1 · p. 4

3.2 Bi-level imitation learning for multi-modal behavior generation

  • The bi-level policy predicts a spatial distribution of short-term goals on a 2D grid, then uses a deterministic goal-conditioned controller to reach a sampled goal.This assigns multimodal behavior modeling to the high-level predictor and lets the low-level policy reuse goal-reaching skills, improving sample efficiency.Source: §3.2 · pp. 4–5

3.3 Prediction and planning for long-horizon stability

  • To stabilize long rollouts beyond the training data’s coverage, the planner samples policy actions and selects among them using predicted future states and rule-based costs.The costs penalize collisions and road departures; future-agent motion prediction is deterministic in this work.Source: §3.3 · p. 5

3.4 Evaluation metrics for traffic simulation

  • Because simulation lacks a single ground truth, the proposed metrics cover rule violations, rollout statistics, and learned likelihoods; coverage and diversity use spatial rollout densities, while a learned occupancy model scores trajectories.Diversity averages pairwise Wasserstein distances between trial density profiles; the learned model predicts future agent positions from real-world logs.Source: §3.4 · p. 6

4 Evaluation

Across Lyft and nuScenes, BITS balances diversity, coverage, and low failure rates; evaluation uses closed-loop vehicle simulations and several realism metrics.

4 Evaluation

  • The study evaluates BITS and its claims on Lyft and nuScenes using a framework that converts dataset scenes into interactive environments.The framework unifies dataset formats, and simulations run on validation scenes unseen during training.Source: §4 · p. 6, §4.1 · p. 7

4.1 Evaluation setup

  • Experiments use 100 validation scenes per dataset, with vehicle-only agents in 10 Hz simulations lasting 20 seconds.Lyft supplies 1,000 hours of auto-labeled data along a 6.8-mile Palo Alto route; nuScenes supplies 5.5 hours of manually labeled data across Boston and Singapore.Source: §4.1 · pp. 6–7
  • The evaluation combines environment measures of failures, coverage, and diversity with dataset-statistic and learned-likelihood measures of realism.Dataset metrics compare speed and jerk histograms and report sADE only as a reference; learned metrics use a likelihood model trained on real driving logs.Source: §4.1 · pp. 7–8

4.2 Main results

  • BITS achieves lower failure rates than recorded trajectories in both datasets, whose ground-truth labels contain noise.The passage reports higher labeling noise in Lyft than nuScenes, with many errors from imprecise vehicle bounding boxes.Source: §4.2 · p. 8
  • On nuScenes, failure rates rise over simulation for non-hierarchical baselines but remain low for BITS and its ablations.Both BITS variants also achieve low failure rates relative to baselines without prediction-and-planning.Source: §4.2 · p. 8
  • BITS combines high diversity and coverage with a low failure rate, unlike baselines that trade diversity against failures.This contrast is stronger on nuScenes, where BITS remains balanced even without prediction-and-planning.Source: §4.2 · pp. 8–9, Figure 2 · p. 8
  • Prediction-and-planning has negligible effects when action samples lack diversity or all lead to failure.The paper exemplifies these cases with TPP on Lyft and SocialGAN on both datasets.Source: §4.2 · p. 9
  • BITS matches SimNet on nuScenes dataset metrics, while its learned likelihood scores are comparable to or higher than other baselines.The learned likelihood falls smoothly as Ornstein-Uhlenbeck perturbation grows; nuScenes BITS scores are on par with recorded trajectories.Source: §4.2 · p. 9, Figure 3 · p. 9

5 Discussion and conclusions

BITS is presented as a data-driven model for diverse, realistic, and stable traffic simulation, alongside an open-source tool for converting dataset scenes into interactive environments. The authors caution that realism metrics are only proxies and that training-data coverage limits the scenarios represented.

Limitations and broader impact

  • Traffic-simulation realism metrics remain an open problem, and the proposed metrics are only proxies; the learned metric may be biased by its model and training data.The evaluation also omits traffic-rule constraints such as correct-side driving and traffic-light compliance.Source: §5 · p. 9
  • Data-driven simulators trained on geographically limited datasets may leave verification pipelines limited to the scenarios represented in those regions.The authors note potential safety concerns when tested vehicles are deployed in less-represented regions.Source: §5 · pp. 9–10

Discussion and conclusions

  • Table 3 is identified as reporting learned likelihood scores for different traffic-simulation methods.The supplied passage identifies the table's scope but gives no scores or method comparison.Source: Table 3 · p. 9

Conclusions

  • BITS generates diverse, realistic, and stable traffic simulations on two real-world AV datasets and includes an open-source tool for creating interactive environments from dataset scenes.The conclusion reports comparisons against competitive baselines but provides no specific comparison values in these passages.Source: §5 · p. 10

6 Appendix

The appendix details BITS's rasterized inputs, learned networks, planning costs, evaluation likelihoods, training setup, and baseline models. Its framework moves from short-horizon goal sampling to goal-conditioned actions and selects among them using predicted neighboring-agent motion and rule-based costs.

6 Appendix

  • The framework maps decision context to short-horizon goals and goal-conditioned action sequences, predicts neighboring agents, then selects actions minimizing a rule-based cost.Decision context concatenates the semantic map and rasterized agent history channel-wise.Source: Figure 4 · p. 13

6.1 Implementation details

  • All three networks use semantic-map and rasterized agent-history channels as input, with a shared ResNet-18 backbone; Table 4 lists key BITS parameters.The semantic map and agent-history maps are concatenated channel-wise.Source: §6.1 · p. 13, Table 4 · p. 14
  • The goal-conditional policy predicts velocity and heading-change controls and forward-integrates them through vehicle dynamics, while the future-state predictor forecasts visible neighboring agents.The policy and predictor share encoder weights and are trained jointly; both use L2 regression against recorded trajectories.Source: §6.1 · p. 14
  • The spatial goal network uses a U-Net-style decoder to predict a 2D goal-likelihood grid, location residuals, and heading, trained with cross-entropy and regression losses.The goal grid uses 0.5-meter pixels; the text reports heading prediction had negligible impact on the goal-conditional policy.Source: §6.1 · p. 14
  • Models are trained for 100,000 iterations, with checkpoints selected using offline validation metrics; training takes 20–30 hours on a single GPU on average.The stated hardware is an NVIDIA Tesla V100 GPU with 32 GB of memory.Source: §6.1 · p. 15
  • Planning combines collision and road-departure costs; the road penalty uses a Manhattan-distance map with vehicle size and orientation accounted for by RoIAlign.For nominal trajectories without collision or road departure, both costs are close to zero, though the sigmoid collision loss may remain nonzero.Source: §6.1 · p. 15
  • The learned-likelihood metric uses an occupancy-based prediction model and averages receding-horizon likelihoods across time and agents.Each time step compares the next 20 rollout steps with predicted occupancy.Source: §6.1 · p. 15

6.2 Baselines

  • The baselines include deterministic behavior cloning (SimNet), conditional GAN generation (SocialGAN), a discrete-latent CVAE (TPP), and an agent-centric isotropic-Gaussian CVAE (TrafficSim).TrafficSim's scene-consistency loss is removed because the setup does not assume control over all agents.Source: §6.2 · pp. 15–16

7 Additional results

The additional results define evaluation measures, examine planning-cost weights and prediction horizons, and show further rollouts from both datasets.

7.1 Complete evaluation metrics

  • Failure metrics report failed-agent rates and collision and off-road causes, alongside collision-type fractions and the share of timesteps spent off-road.Off-road failure requires leaving the drivable region for more than one second; collision types are calculated independently, so an agent can have multiple types in an episode.Source: §7.1 · p. 16
  • Coverage and diversity measure rollout variation, while dataset metrics compare simulated and recorded speed, acceleration, and jerk histograms using normalized Wasserstein distances.Coverage is discretized at 2m by 2m over five trials; the driving-profile histograms use 20 bins.Source: §7.1 · p. 16
  • sADE and sFDE measure average and final position differences, but are included only as references because alternative simulations can be heavily penalized.The metrics are not suitable for measuring simulation realism.Source: §7.1 · pp. 16–17

7.2 Planning cost weight ablation

  • Both planning cost terms reduce failures of their corresponding types, with diminishing reductions above certain weights and overall performance insensitive to the weight ratio.The ablation varies collision cost logarithmically and sets off-road cost to either 0.0 or 1.0.Source: §7.2 · pp. 17–18, Figure 6 · p. 18

7.3 Prediction and planning horizon ablation

  • Longer planning horizons H = {50, 80} increase failure rates, while H = {10, 20} perform similarly; shorter horizons improve simulation success but increase acceleration and jerk.The authors attribute long-horizon errors to difficult future prediction under stochastic agent policies, and suggest dynamically adjusting the horizon based on prediction uncertainty.Source: Table 7 · p. 18, §7.3 · p. 18

7.4 Additional qualitative results

  • Additional qualitative rollouts show different simulations from the same initial scene configuration for nuScenes and Lyft, with recorded trajectories in the rightmost column.The nuScenes rollouts appear in Figure 7 and Lyft rollouts in Figure 8.Source: §7.4 · p. 18, Figure 7 · p. 19, Figure 8 · p. 19
Loading 2208.12403v1…