Recent efforts recognize the power of scale in 3D learning (e.g. PTv3) and attention mechanisms (e.g. FlashAt-tention). However, current point cloud backbones fail to holistically unify geometric locality, attention mechanisms, and GPU architectures in one view. In this paper, we introduce Flash3D Transformer, which aligns geometric locality and GPU tiling through a principled locality mechanism based on Perfect Spatial Hashing (PSH). The common alignment with GPU tiling naturally fuses our PSH locality mechanism with FlashAttention at negligible extra cost. This mechanism affords flexible design choices throughout the backbone that result in superior downstream task results. Flash3D outperforms state-of-the-art PTv3 results on benchmark datasets, delivering a 2.25x speed increase and 2.4x memory efficiency boost. This efficiency enables scaling to wider attention scopes and larger models without additional overhead. Such scaling allows Flash3D to achieve even higher task accuracies than PTv3 under the same compute budget.

Simulation is critical for safety evaluation in autonomous driving, particularly in capturing complex interactive behaviors. However, generating realistic and controllable traffic scenarios in long-tail situations remains a significant challenge. Existing generative models suffer from the conflicting objective between user-defined controllability and realism constraints, which is amplified in safety-critical contexts. In this work, we introduce the Causal Compositional Diffusion Model (CCDiff), a structure-guided diffusion framework to address these challenges. We first formulate the learning of controllable and realistic closed-loop simulation as a constrained optimization problem. Then, CCDiff maximizes controllability while adhering to realism by automatically identifying and injecting causal structures directly into the diffusion process, providing structured guidance to enhance both realism and controllability. Through rigorous evaluations on benchmark datasets and in a closed-loop simulator, CCDiff demonstrates substantial gains over state-of-the-art approaches in generating realistic and user-preferred trajectories. Our results show CCDiff’s effectiveness in extracting and leveraging causal structures, showing improved closed-loop performance based on key metrics such as collision rate, off-road rate, FDE, and comfort. For more details, welcome to check our project website.

It is difficult to anticipate the myriad challenges that a predictive model will encounter once deployed. Common practice entails a reactive, cyclical approach: model deployment, data mining, and retraining. We instead develop a proactive longtail discovery process by imagining additional data during training. In particular, we develop general model-based longtail signals, including a differentiable, single forward pass formulation of epistemic uncertainty that does not impact model parameters or predictive performance but can flag rare or hard inputs. We leverage these signals as guidance to generate additional training data from a latent diffusion model in a process we call Longtail Guidance (LTG). Crucially, we can perform LTG without retraining the diffusion model or the predictive model, and we do not need to expose the predictive model to intermediate diffusion states. Data generated by LTG exhibit semantically meaningful variation, yield significant generalization improvements on numerous image classification benchmarks, and can be analyzed by a VLM to proactively discover, textually explain, and address conceptual gaps in a deployed predictive model.

We present DriveGPT, a scalable behavior model for autonomous driving. We model driving as a sequential decision-making task, and learn a transformer model to predict future agent states as tokens in an autoregressive fashion. We scale up our model parameters and training data by multiple orders of magnitude, enabling us to explore the scaling properties in terms of dataset size, model parameters, and compute. We evaluate DriveGPT across different scales in a planning task, through both quantitative metrics and qualitative examples, including closed-loop driving in complex real-world scenarios. In a separate prediction task, DriveGPT outperforms state-of-the-art baselines and exhibits improved performance by pretraining on a large-scale dataset, further validating the benefits of data scaling.

Multi-frame temporal inputs are important for vision-based autonomous driving. Observations from different angles enable the recovery of 3 D object states from 2 D images as long as we can identify the same instance from different input frames. However, the dynamic nature of driving scenes leads to significant variance in the instance appearance and shape captured by the cameras at different time steps. To this end, we propose a novel contrastive learning algorithm, Cohere3D, to learn coherent instance representations robust to the changes of distance and perspective in a long-term temporal sequence without any human annotations. In the pretraining stage, raw point clouds from LiDAR sensors are utilized to construct the instance-wise long-term temporal correspondence, which serves as guidance for the extraction of instance-level representation from the vision-based bird's-eye-view (BEV) feature map. Cohere3D encourages consistent representation for the same instance at different frames but distinguishes between different instances. We validate the effectiveness and generalizability of our algorithm by finetuning the pretrained model across key downstream autonomous driving tasks: perception, mapping, prediction, and planning. Results show a notable improvement in both data efficiency and final performance in all these tasks.

In this paper, we introduce the first published planner to drive a car in dense, urban traffic using Inverse Reinforcement Learning (IRL). Our planner, DriveIRL, generates a diverse set of trajectory proposals and scores them with a learned model. The best trajectory is tracked by our self-driving vehicle's low-level controller. We train our trajectory scoring model on a 500+ hour real-world dataset of expert driving demonstrations in Las Vegas within the maximum entropy IRL framework. DriveIRL's benefits include: a simple design due to only learning the trajectory scoring function, a flexible and relatively interpretable feature engineering approach, and strong real-world performance. We validated DriveIRL on the Las Vegas Strip and demonstrated fully autonomous driving in heavy traffic, including scenarios involving cut-ins, abrupt braking by the lead vehicle, and hotel pickup/dropoff zones. Our dataset, a part of nuPlan, has been released to the public to help further research in this area.

We present CoverNet, a new method for multimodal, probabilistic trajectory prediction for urban driving. Previous work has employed a variety of methods, including multimodal regression, occupancy maps, and 1-step stochastic policies. We instead frame the trajectory prediction problem as classification over a diverse set of trajectories. The size of this set remains manageable due to the limited number of distinct actions that can be taken over a reasonable prediction horizon. We structure the trajectory set to a) ensure a desired level of coverage of the state space, and b) eliminate physically impossible trajectories. By dynamically generating trajectory sets based on the agent's current state, we can further improve our method's efficiency. We demonstrate our approach on public, real world self-driving datasets, and show that it outperforms state-of-the-art methods.

This paper presents a method for optimal trajectory generation for discrete-time nonlinear systems with linear temporal logic (LTL) task specifications. Our approach is based on recent advances in stochastic optimization algorithms for optimal trajectory generation. These methods rely on estimation of the rare event of sampling optimal trajectories, which is achieved by incrementally improving a sampling distribution so as to minimize the cross-entropy. A key component of these stochastic optimization algorithms is determining whether or not a trajectory is collision-free. We generalize this collision checking to efficiently verify whether or not a trajectory satisfies a LTL formula. Interestingly, this verification can be done in time polynomial in the length of the LTL formula and the trajectory. We also propose a method for efficiently re-using parts of trajectories that only partially satisfy the specification, instead of simply discarding the entire sample. Our approach is demonstrated through numerical experiments involving Dubins car and a generic point-mass model subject to complex temporal logic task specifications.

We present a mathematical programming-based method for optimal control of discrete-time dynamical systems subject to temporal logic task specifications. We use linear temporal logic (LTL) to specify a wide range of properties and tasks, such as safety, progress, response, surveillance, repeated assembly, and environmental monitoring. Our method directly encodes an LTL formula as mixed-integer linear constraints on the continuous system variables, avoiding the computationally expensive processes of creating a finite abstraction of the system and a Büchi automaton for the specification. In numerical experiments, we solve temporal logic motion planning tasks for high-dimensional (10+ continuous state) dynamical systems.

We introduce an algorithm for the optimal control of stochastic nonlinear systems subject to temporal logic constraints on their behavior. We compute directly on the state space of the system, avoiding the expensive pre-computation of a discrete abstraction. An automaton that corresponds to the temporal logic specification guides the computation of a control policy that maximizes the probability that the system satisfies the specification. This reduces controller synthesis to solving a sequence of stochastic constrained reachability problems. Each individual reachability problem is solved via the Hamilton-Jacobi-Bellman (HJB) partial differential equation of stochastic optimal control theory. To increase the efficiency of our approach, we exploit a class of systems where the HJB equation is linear due to structural assumptions on the noise. The linearity of the partial differential equation allows us to pre-compute control policy primitives and then compose them, at essentially zero cost, to conservatively satisfy a complex temporal logic specification.

Motivated by robotic motion planning, we develop a framework for control policy synthesis for both non-deterministic transition systems and Markov decision processes that are subject to temporal logic task specifications. We introduce a fragment of linear temporal logic that can be used to specify common motion planning tasks such as safe navigation, response to the environment, persistent coverage, and surveillance. This fragment is computationally efficient; the complexity of control policy synthesis is a doubly-exponential improvement over standard linear temporal logic for both non-deterministic transition systems and Markov decision processes. This improvement is possible because we compute directly on the original system, as opposed to the automata-based approach commonly used. We give simulation results for representative motion planning tasks and compare to generalized reactivity(1).

We develop a method for the control of discrete-time nonlinear systems subject to temporal logic specifications. Our approach uses a coarse abstraction of the system and an automaton representing the temporal logic specification to guide the search for a feasible trajectory. This decomposes the search for a feasible trajectory into a series of constrained reachability problems. Thus, one can create controllers for any system for which techniques exist to compute (approximate) solutions to constrained reachability problems. Representative techniques include sampling-based methods for motion planning, reachable set computations for linear systems, and graph search for finite discrete systems. Our approach avoids the expensive computation of a discrete abstraction, and its implementation is amenable to parallel computing. We demonstrate our approach with numerical experiments on temporal logic motion planning problems with high-dimensional (10+ states) continuous systems.