Traffic simulation aims to learn a policy for traffic agents that, when unrolled in closed-loop, faithfully recovers the joint distribution of trajectories observed in the real world. Inspired by large language models, tokenized multi-agent policies have recently become the state-of-the-art in traffic simulation. However, they are typically trained through open-loop behavior cloning, and thus suffer from covariate shift when executed in closed-loop during simulation. In this work, we present Closest Among Top-K (CAT-K) rollouts, a simple yet effective closed-loop fine-tuning strategy to mitigate covariate shift. CAT-K fine-tuning only requires existing trajectory data, without reinforcement learning or generative adversarial imitation. Concretely, CAT-K fine-tuning enables a small 7M-parameter tokenized traffic simulation policy to outperform a 102M-parameter model from the same model family, achieving the top spot on the Waymo Sim Agent Challenge leaderboard at the time of submission. The code is available at https://github.com/NVlabs/catk.

We present STORM, a spatio-temporal reconstruction model designed for reconstructing dynamic outdoor scenes from sparse observations. Existing dynamic reconstruction methods often rely on per-scene optimization, dense observations across space and time, and strong motion supervision, resulting in lengthy optimization times, limited generalization to novel views or scenes, and degenerated quality caused by noisy pseudo-labels for dynamics. To address these challenges, STORM leverages a data-driven Transformer architecture that directly infers dynamic 3D scene representations—parameterized by 3D Gaussians and their velocities—in a single forward pass. Our key design is to aggregate 3D Gaussians from all frames using self-supervised scene flows, transforming them to the target timestep to enable complete (i.e., "amodal") reconstructions from arbitrary viewpoints at any moment in time. As an emergent property, STORM automatically captures dynamic instances and generates high-quality masks using only reconstruction losses. Extensive experiments on public datasets show that STORM achieves precise dynamic scene reconstruction, surpassing state-of-the-art per-scene optimization methods (+4.3 to 6.6 PSNR) and existing feed-forward approaches (+2.1 to 4.7 PSNR) in dynamic regions. STORM reconstructs large-scale outdoor scenes in 200ms, supports real-time rendering, and outperforms competitors in scene flow estimation, improving 3D EPE by 0.422m and Acc5 by 28.02%. Beyond reconstruction, we showcase four additional applications of our model, illustrating the potential of self-supervised learning for broader dynamic scene understanding. For more details, please visit our project at https://jiawei-yang.github.io/STORM/.

Adept traffic models are critical to both real-time prediction/planning and closed-loop simulation for autonomous vehicles (AV). Key design objectives include accuracy, diverse multimodal behaviors, interpretability, and compatibility with other modules in the autonomy stack, e.g., the downstream planner. We present Categorical Traffic Transformer (CTT), a traffic model that outputs both continuous trajectory predictions and categorical predictions with clear semantic meanings (lane modes, homotopies, etc.). The most outstanding feature of CTT is its fully interpretable latent space, which enables direct supervision of the latent variables from the ground truth during training and avoids mode collapse completely. As a result, CTT can generate diverse behaviors conditioned on different semantic modes while significantly beating SOTA on prediction accuracy. In addition, CTT's ability to input and output tokens enables direct integration with semantic-heavy modules such as behavior planners and language models, bridging the tokenized representation and the continuous trajectory space.

Autonomous driving necessitates the ability to reason about future interactions between traffic agents and to make informed evaluations for planning. This paper introduces the Gen-Drive framework, which shifts from the traditional prediction and deterministic planning framework to a generation-then-evaluation planning paradigm. The framework employs a behavior diffusion model as a scene generator to produce diverse possible future scenarios, thereby enhancing the capability for joint interaction reasoning. To facilitate decision-making, we propose a scene evaluator (reward) model, trained with pairwise preference data collected through VLM assistance, thereby reducing human workload and enhancing scalability. Furthermore, we utilize an RL fine-tuning framework to improve the generation quality of the diffusion model, rendering it more effective for planning tasks. We conduct training and closed-loop planning tests on the nuPlan dataset, and the results demonstrate that employing such a generation-then-evaluation strategy outperforms other learning-based approaches. Additionally, the fine-tuned generative driving policy shows significant enhancements in planning performance. We further demonstrate that utilizing our learned reward model for evaluation or RL fine-tuning leads to better planning performance compared to relying on human-designed rewards. Project website: https://mczhi.github.io/GenDrive.

Synthesizing photo-realistic visual observations from an ego vehicle's driving trajectory is a critical step towards scalable training of self-driving models. Reconstruction-based methods create 3D scenes from driving logs and synthesize geometry-consistent driving videos through neural rendering, but their dependence on costly object annotations limits their ability to generalize to in-the-wild driving scenarios. On the other hand, generative models can synthesize action-conditioned driving videos in a more generalizable way but often struggle with maintaining 3D visual consistency. In this paper, we present DreamDrive, a 4D spatial-temporal scene generation approach that combines the merits of generation and reconstruction, to synthesize generalizable 4D driving scenes and dynamic driving videos with 3D consistency. Specifically, we leverage the generative power of video diffusion models to synthesize a sequence of visual references and further elevate them to 4D with a novel hybrid Gaussian representation. Given a driving trajectory, we then render 3D-consistent driving videos via Gaussian splatting. The use of generative priors allows our method to produce high-quality 4D scenes from in-the-wild driving data, while neural rendering ensures 3D-consistent video generation from the 4D scenes. Extensive experiments on nuScenes and in-the-wild driving data demonstrate that DreamDrive can generate controllable and generalizable 4D driving scenes, synthesize novel views of driving videos with high fidelity and 3D consistency, decompose static and dynamic elements in a self-supervised manner, and enhance perception and planning tasks for autonomous driving.

Advancements in deep multi-agent reinforcement learning (MARL) have positioned it as a promising approach for decision-making in cooperative games. However, it still remains challenging for MARL agents to learn cooperative strategies for some game environments. Recently, large language models (LLMs) have demonstrated emergent reasoning capabilities, making them promising candidates for enhancing coordination among the agents. However, due to the model size of LLMs, it can be expensive to frequently infer LLMs for actions that agents can take. In this work, we propose You Only LLM Once for MARL (YOLO-MARL), a novel framework that leverages the high-level task planning capabilities of LLMs to improve the policy learning process of multi-agents in cooperative games. Notably, for each game environment, YOLO-MARL only requires one time interaction with LLMs in the proposed strategy generation, state interpretation and planning function generation modules, before the MARL policy training process. This avoids the ongoing costs and computational time associated with frequent LLMs API calls during training. Moreover, trained decentralized policies based on normal-sized neural networks operate independently of the LLM. We evaluate our method across two different environments and demonstrate that YOLO-MARL outperforms traditional MARL algorithms. The Github repository of our code can be found at https://github.com/paulzyzy/YOLO-MARL.

Motion prediction and cost evaluation are vital components in the decision-making system of autonomous vehicles. However, existing methods often ignore the importance of cost learning and treat them as separate modules. In this study, we employ a tree-structured policy planner and propose a differentiable joint training framework for both ego-conditioned prediction and cost models, resulting in a direct improvement of the final planning performance. For conditional prediction, we introduce a query-centric Transformer model that performs efficient ego-conditioned motion prediction. For planning cost, we propose a learnable context-aware cost function with latent interaction features, facilitating differentiable joint learning. We validate our proposed approach using the real-world nuPlan dataset and its associated planning test platform. Our framework not only matches state-of-the-art planning methods but outperforms other learning-based methods in planning quality, while operating more efficiently in terms of runtime. We show that joint training delivers significantly better performance than separate training of the two modules. Additionally, we find that tree-structured policy planning outperforms the conventional single-stage planning approach. Code is available: https://github.com/MCZhi/DTPP.

Autonomous vehicles (AVs) need to reason about the multimodal behavior of neighboring agents while planning their own motion. Many existing trajectory planners seek a single trajectory that performs well under all plausible futures simultaneously, ignoring bi-directional interactions and thus leading to overly conservative plans. Policy planning, whereby the ego agent plans a policy that reacts to the environment's multimodal behavior, is a promising direction as it can account for the action-reaction interactions between the AV and the environment. However, most existing policy planners do not scale to the complexity of real autonomous vehicle applications: they are either not compatible with modern deep learning prediction models, not interpretable, or not able to generate high quality trajectories. To fill this gap, we propose Tree Policy Planning (TPP), a policy planner that is compatible with state-of-the-art deep learning prediction models, generates multistage motion plans, and accounts for the influence of ego agent on the environment behavior. The key idea of TPP is to reduce the continuous optimization problem into a tractable discrete Markov Decision Process (MDP) through the construction of two tree structures: an ego trajectory tree for ego trajectory options, and a scenario tree for multi-modal ego-conditioned environment predictions. We demonstrate the efficacy of TPP in closed-loop simulations based on real-world nuScenes dataset and results show that TPP scales to realistic AV scenarios and significantly outperforms non-policy baselines.

Drivers have a responsibility to exercise reasonable care to avoid collision with other road users. This assumed responsibility allows interacting agents to maintain safety without explicit coordination. Thus to enable safe autonomous vehicle (AV) interactions, AVs must understand what their responsibilities are to maintain safety and how they affect the safety of nearby agents. In this work we seek to understand how responsibility is shared in multi-agent settings where an autonomous agent is interacting with human counterparts. We introduce Responsibility-Aware Control Barrier Functions (RA-CBFs) and present a method to learn responsibility allocations from data. By combining safety-critical control and learning- based techniques, RA-CBFs allow us to account for scene- dependent responsibility allocations and synthesize safe and efficient driving behaviors without making worst-case assumptions that typically result in overly-conservative behaviors. We test our framework using real-world driving data and demonstrate its efficacy as a tool for both safe control and forensic analysis of unsafe driving.

Autonomous vehicles must often contend with conflicting planning requirements, e.g., safety and comfort could be at odds with each other if avoiding a collision calls for slamming the brakes. To resolve such conflicts, assigning importance ranking to rules (i.e., imposing a rule hierarchy) has been proposed, which, in turn, induces rankings on trajectories based on the importance of the rules they satisfy. On one hand, imposing rule hierarchies can enhance interpretability, but introduce combinatorial complexity to planning; while on the other hand, differentiable reward structures can be leveraged by modern gradient-based optimization tools, but are less interpretable and unintuitive to tune. In this paper, we present an approach to equivalently express rule hierar-chies as differentiable reward structures amenable to modern gradient-based optimizers, thereby, achieving the best of both worlds. We achieve this by formulating rank-preserving reward functions that are monotonic in the rank of the trajectories induced by the rule hierarchy; i.e., higher ranked trajectories receive higher reward. Equipped with a rule hierarchy and its corresponding rank-preserving reward function, we develop a two-stage planner that can efficiently resolve conflicting planning requirements. We demonstrate that our approach can generate motion plans in ~7-10 Hz for various challenging road navigation and intersection negotiation scenarios.

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 devising first principle models for human-like behaviors is generally infeasible. In this work, we take a data-driven approach 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 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 with scenarios from two large-scale driving datasets and show our method 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 video results and code release, see https://bit.ly/3L9uzj3.

Controllable and realistic traffic simulation is critical for developing and verifying autonomous vehicles. Typical heuristic-based traffic models offer flexible control to make vehicles follow specific trajectories and traffic rules. On the other hand, data-driven approaches generate realistic and human-like behaviors, improving transfer from simulated to real-world traffic. However, to the best of our knowledge, no traffic model offers both controllability and realism. In this work, we develop a conditional diffusion model for controllable traffic generation (CTG) that allows users to control desired properties of trajectories at test time (e.g., reach a goal or follow a speed limit) while maintaining realism and physical feasibility through enforced dynamics. The key technical idea is to leverage recent advances from diffusion modeling and differentiable logic to guide generated trajectories to meet rules defined using signal temporal logic (STL). We further extend guidance to multi-agent settings and enable interaction-based rules like collision avoidance. CTG is extensively evaluated on the nuScenes dataset for diverse and composite rules, demonstrating improvement over strong baselines in terms of the controllability-realism tradeoff. Demo videos can be found at https://aiasd.github.io/ctg.github.io

The ability to anticipate surrounding agents' behaviors is critical to enable safe and seamless autonomous vehicles (AVs). While phenomenological methods have successfully predicted future trajectories from scene context, these predictions lack interpretability. On the other hand, ontological approaches assume an underlying structure able to describe the interaction dynamics or agents' internal decision processes. Still, they often suffer from poor scalability or cannot reflect diverse human behaviors. This work proposes an interpretability framework for a phenomenological method through responsibility evaluations. We formulate responsibility as a measure of how much an agent takes into account the welfare of other agents through counterfactual reasoning. Additionally, this framework abstracts the computed responsibility sequences into different responsibility levels and grounds these latent levels into reward functions. The proposed responsibility-based interpretability framework is modular and easily integrated into a wide range of prediction models. To demonstrate the utility of the proposed framework in providing added interpretability, we adapt an existing AV prediction model and perform a simulation study on a real-world nuScenes traffic dataset. Experimental results show that we can perform offline ex-post traffic analysis by incorporating the responsibility signal and rendering interpretable but accurate online trajectory predictions.

A critical task for developing safe autonomous driving stacks is to determine whether an obstacle is safety-critical, i.e., poses an imminent threat to the autonomous vehicle. Our previous work showed that Hamilton Jacobi reachability theory can be applied to compute interaction-dynamics-aware perception safety zones that better inform an ego vehicle’s perception module which obstacles are considered safety-critical. For completeness, these zones are typically larger than absolutely necessary, forcing the perception module to pay attention to a larger collection of objects for the sake of conservatism. As an improvement, we propose a maneuver-based decomposition of our safety zones that leverages information about the ego maneuver to reduce the zone volume. In particular, we propose a "temporal convolution" operation that produces safety zones for specific ego maneuvers, thus limiting the ego’s behavior to reduce the size of the safety zones. We show with numerical experiments that maneuver-based zones are significantly smaller (up to 76% size reduction) than the baseline while maintaining completeness.

Trajectory prediction is a critical functionality of autonomous systems that share environments with uncontrolled agents, one prominent example being self-driving vehicles. Currently, most prediction methods do not enforce scene consistency, i.e., there are a substantial amount of self-collisions between predicted trajectories of different agents in the scene. Moreover, many approaches generate individual trajectory predictions per agent instead of joint trajectory predictions of the whole scene, which makes downstream planning difficult. In this work, we present ScePT, a policy planning-based trajectory prediction model that generates accurate, scene-consistent trajectory predictions suitable for autonomous system motion planning. It explicitly enforces scene consistency and learns an agent interaction policy that can be used for conditional prediction. Experiments on multiple real-world pedestrians and autonomous vehicle datasets show that ScePT matches current state-of-the-art prediction accuracy with significantly improved scene consistency. We also demonstrate ScePT's ability to work with a downstream contingency planner.

To enable safe autonomous vehicle (AV) operations, it is critical that an AV’s obstacle detection module can reliably detect obstacles that pose a safety threat (i.e., are safety-critical). It is therefore desirable that the evaluation metric for the perception system captures the safety-criticality of objects. Unfortunately, existing perception evaluation metrics tend to make strong assumptions about the objects and ignore the dynamic interactions between agents, and thus do not accurately capture the safety risks in reality. To address these shortcomings, we introduce an interaction-dynamics-aware obstacle detection evaluation metric by accounting for closed-loop dynamic interactions between an ego vehicle and obstacles in the scene. By borrowing existing theory from optimal control theory, namely Hamilton-Jacobi reachability, we present a computationally tractable method for constructing a “safety zone a region in state space that defines where safety-critical obstacles lie for the purpose of defining safety metrics. Our proposed safety zone is mathematically complete, and can be easily computed to reflect a variety of safety requirements. Using an off-the-shelf detection algorithm from the nuScenes detection challenge leaderboard, we demonstrate that our approach is computationally lightweight, and can better capture safety-critical perception errors than a baseline approach.