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.

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.

Sequentially solving similar optimization problems under strict runtime constraints is essential for many applications, such as robot control, autonomous driving, and portfolio management. The performance of local optimization methods in these settings is sensitive to the initial solution: poor initialization can lead to slow convergence or suboptimal solutions. To address this challenge, we propose learning to predict \emph{multiple} diverse initial solutions given parameters that define the problem instance. We introduce two strategies for utilizing multiple initial solutions: (i) a single-optimizer approach, where the most promising initial solution is chosen using a selection function, and (ii) a multiple-optimizers approach, where several optimizers, potentially run in parallel, are each initialized with a different solution, with the best solution chosen afterward. We validate our method on three optimal control benchmark tasks: cart-pole, reacher, and autonomous driving, using different optimizers: DDP, MPPI, and iLQR. We find significant and consistent improvement with our method across all evaluation settings and demonstrate that it efficiently scales with the number of initial solutions required. The code is available at $\href{https://github.com/EladSharony/miso}{\tt{https://github.com/EladSharony/miso}}$.

Distribution shifts between operational domains can severely affect the performance of learned models in self-driving vehicles (SDVs). While this is a well-established problem, prior work has mostly explored naive solutions such as fine-tuning, focusing on the motion prediction task. In this work, we explore novel adaptation strategies for differentiable autonomy stacks consisting of prediction, planning, and control, perform evaluation in closed-loop, and investigate the often-overlooked issue of catastrophic forgetting. Specifically, we introduce two simple yet effective techniques: a low-rank residual decoder (LoRD) and multi-task fine-tuning. Through experiments across three models conducted on two real-world autonomous driving datasets (nuPlan, exiD), we demonstrate the effectiveness of our methods and highlight a significant performance gap between open-loop and closed-loop evaluation in prior approaches. Our approach improves forgetting by up to 23.33% and the closed-loop OOD driving score by 8.83% in comparison to standard fine-tuning.

We propose DistillNeRF, a self-supervised learning framework addressing the challenge of understanding 3D environments from limited 2D observations in autonomous driving. Our method is a generalizable feedforward model that predicts a rich neural scene representation from sparse, single-frame multi-view camera inputs, and is trained self-supervised with differentiable rendering to reconstruct RGB, depth, or feature images. Our first insight is to exploit per-scene optimized Neural Radiance Fields (NeRFs) by generating dense depth and virtual camera targets for training, thereby helping our model to learn 3D geometry from sparse non-overlapping image inputs. Second, to learn a semantically rich 3D representation, we propose distilling features from pre-trained 2D foundation models, such as CLIP or DINOv2, thereby enabling various downstream tasks without the need for costly 3D human annotations. To leverage these two insights, we introduce a novel model architecture with a two-stage lift-splat-shoot encoder and a parameterized sparse hierarchical voxel representation. Experimental results on the NuScenes dataset demonstrate that DistillNeRF significantly outperforms existing comparable self-supervised methods for scene reconstruction, novel view synthesis, and depth estimation; and it allows for competitive zero-shot 3D semantic occupancy prediction, as well as open-world scene understanding through distilled foundation model features. Demos and code will be available at https://distillnerf.github.io/.

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 hierarchies 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.

In highly interactive driving scenarios, the actions of one agent greatly influences those of its neighbors. Planning safe motions for autonomous vehicles in such interactive environments, therefore, requires reasoning about the impact of the ego's intended motion plan on nearby agents' behavior. Deep-learning-based models have recently achieved great success in trajectory prediction and many models in the literature allow for ego-conditioned prediction. However, leveraging ego-conditioned prediction remains challenging in downstream planning due to the complex nature of neural networks, limiting the planner structure to simple ones, e.g., sampling-based planner. Despite their ability to generate fine-grained high-quality motion plans, it is difficult for gradient-based planning algorithms, such as model predictive control (MPC), to leverage ego-conditioned prediction due to their iterative nature and need for gradient. We present Interactive Joint Planning (IJP) that bridges MPC with learned prediction models in a computationally scalable manner to provide us the best of both the worlds. In particular, IJP jointly optimizes over the behavior of the ego and the surrounding agents and leverages deep-learned prediction models as prediction priors that the join trajectory optimization tries to stay close to. Furthermore, by leveraging homotopy classes, our joint optimizer searches over diverse motion plans to avoid getting stuck at local minima. Closed-loop simulation result shows that IJP significantly outperforms the baselines that are either without joint optimization or running sampling-based planning.

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.

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 \emph{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.

Intelligent robotic systems, such as autonomous vehicles (AVs), are typically architected in a modular fashion and comprised of modules performing detection, tracking, prediction, planning, and control, among others [1, 2, 3, 4, 5, 6, 7, 8]. Modular architectures are generally desirable because of their verifiability, interpretability and generalization performance; however, they also suffer from compounding errors, information bottlenecks, and integration challenges. A promising line of work tackling these issues focuses on making AV stacks more integrated (by relaxing inter-module interfaces) and data-driven (by optimizing modules jointly with respect to their downstream task). For example, in the context of AV perception, recent work has achieved substantial performance gains by jointly training tracking models with detection [9] and prediction models [10, 11]. To extend such a joint, data-driven approach to decision making, existing approaches replace hand-engineered components, e.g., planning and control algorithms, with deep neural networks [12, 13, 14]. As neural networks are differentiable, they can be optimized end-to-end for a final control objective; however, they offer weaker generalization, little to no interpretability or safety guarantees. We introduce DiffStack, a differentiable AV stack with modules for prediction, planning, and control that combines the benefits of modular and data-driven architectures (Figure 1). The prediction module in DiffStack is a learned neural network that predicts the future motion of agents; the planning and control modules are principled, hand-engineered algorithms that produce AV actions given the current world state and motion predictions. Importantly, our hand-engineered planning and control algorithms are differentiable, enabling the training of the upstream prediction module for a downstream control objective by backpropagating gradients through the algorithms.