Realistic and interactive scene simulation is a key prerequisite for autonomous vehicle (AV) development. In this work, we present SceneDiffuser, a scene-level diffusion prior designed for traffic simulation. It offers a unified framework that addresses two key stages of simulation: scene initialization, which involves generating initial traffic layouts, and scene rollout, which encompasses the closed-loop simulation of agent behaviors. While diffusion models have been proven effective in learning realistic and multimodal agent distributions, several challenges remain, including controllability, maintaining realism in closed-loop simulations, and ensuring inference efficiency. To address these issues, we introduce amortized diffusion for simulation. This novel diffusion denoising paradigm amortizes the computational cost of denoising over future simulation steps, significantly reducing the cost per rollout step (16x less inference steps) while also mitigating closed-loop errors. We further enhance controllability through the introduction of generalized hard constraints, a simple yet effective inference-time constraint mechanism, as well as language-based constrained scene generation via few-shot prompting of a large language model (LLM). Our investigations into model scaling reveal that increased computational resources significantly improve overall simulation realism. We demonstrate the effectiveness of our approach on the Waymo Open Sim Agents Challenge, achieving top open-loop performance and the best closed-loop performance among diffusion models.

LMMs have shown impressive visual understanding capabilities, with the potential to be applied in agents, which demand strong reasoning and planning abilities. Nevertheless, existing benchmarks mostly assess their reasoning abilities in language part, where the chain-of-thought is entirely composed of text.We consider the scenario where visual signals are continuously updated and required along the decision making process. Such vision-in-the-chain reasoning paradigm is more aligned with the needs of multimodal agents, while being rarely evaluated. In this paper, we introduce MageBench, a reasoning capability oriented multimodal agent benchmark that, while having light-weight environments, poses significant reasoning challenges and holds substantial practical value. This benchmark currently includes three types of environments: WebUI, Sokoban, and Football, comprising a total of 483 different scenarios. It thoroughly validates the agent's knowledge and engineering capabilities, visual intelligence, and interaction skills. The results show that only a few product-level models are better than random acting, and all of them are far inferior to human-level. More specifically, we found current models severely lack the ability to modify their planning based on visual feedback, as well as visual imagination, interleaved image-text long context handling, and other abilities. We hope that our work will provide optimization directions for LMM from the perspective of being an agent. We release our code and data at https://github.com/microsoft/MageBench.

Building generalist agents that can rapidly adapt to new environments is a key challenge for deploying AI in the digital and real worlds. Is scaling current agent architectures the most effective way to build generalist agents? We propose a novel approach to pre-train relatively small policies on relatively small datasets and adapt them to unseen environments via in-context learning, without any finetuning. Our key idea is that retrieval offers a powerful bias for fast adaptation. Indeed, we demonstrate that even a simple retrieval-based 1-nearest neighbor agent offers a surprisingly strong baseline for today's state-of-the-art generalist agents. From this starting point, we construct a semi-parametric agent, REGENT, that trains a transformer-based policy on sequences of queries and retrieved neighbors. REGENT can generalize to unseen robotics and game-playing environments via retrieval augmentation and in-context learning, achieving this with up to 3x fewer parameters and up to an order-of-magnitude fewer pre-training datapoints, significantly outperforming today's state-of-the-art generalist agents. AI agents, both in the digital [38, 19, 37, 28, 53] and real world [5, 7, 63, 33, 48, 24], constantly face changing environments that require rapid or even instantaneous adaptation. True generalist agents must not only be capable of performing well on large numbers of training environments, but arguably more importantly, they must be capable of adapting rapidly to new environments. While this goal has been of considerable interest to the reinforcement learning research community, it has proven elusive. The most promising results so far have all been attributed to large policies [38, 19, 37, 28, 5], pre-trained on large datasets across many environments, and even these models still struggle to generalize to unseen environments without many new environment-specific demonstrations. In this work, we take a different approach to the problem of constructing such generalist agents. We start by asking: Is scaling current agent architectures the most effective way to build generalist agents? Observing that retrieval offers a powerful bias for fast adaptation, we first evaluate a simple 1-nearest neighbor method: "Retrieve and Play (R&P)". To determine the action at the current state, R&P simply retrieves the closest state from a few demonstrations in the target environment and plays its corresponding action. Tested on a wide range of environments, both robotics and game-playing, R&P performs on-par or better than the state-of-the-art generalist agents.

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.

It is significant to employ multiple autonomous underwater vehicles (AUVs) to execute the underwater target tracking task collaboratively. However, it's pretty challenging to meet various prerequisites utilizing traditional control methods. Therefore, we propose an effective two-stage learning from demonstrations training framework, FISHER, to highlight the adaptability of reinforcement learning (RL) methods in the multi-AUV underwater target tracking task, while addressing its limitations such as extensive requirements for environmental interactions and the challenges in designing reward functions. The first stage utilizes imitation learning (IL) to realize policy improvement and generate offline datasets. To be specific, we introduce multi-agent discriminator-actor-critic based on improvements of the generative adversarial IL algorithm and multi-agent IL optimization objective derived from the Nash equilibrium condition. Then in the second stage, we develop multi-agent independent generalized decision transformer, which analyzes the latent representation to match the future states of high-quality samples rather than reward function, attaining further enhanced policies capable of handling various scenarios. Besides, we propose a simulation to simulation demonstration generation procedure to facilitate the generation of expert demonstrations in underwater environments, which capitalizes on traditional control methods and can easily accomplish the domain transfer to obtain demonstrations. Extensive simulation experiments from multiple scenarios showcase that FISHER possesses strong stability, multi-task performance and capability of generalization.

Robots in densely populated real-world environments frequently encounter constrained and cluttered situations such as passing through narrow doorways, hallways, and corridor intersections, where conflicts over limited space result in collisions or deadlocks among the robots. Current decentralized state-of-the-art optimization- and neural network-based approaches (i) are predominantly designed for general open spaces, and (ii) are overly conservative, either guaranteeing safety, or liveness, but not both. While some solutions rely on centralized conflict resolution, their highly invasive trajectories make them impractical for real-world deployment. This paper introduces LiveNet, a fully decentralized and robust neural network controller that enables human-like yielding and passing, resulting in agile, non-conservative, deadlock-free, and safe, navigation in congested, conflict-prone spaces. LiveNet is minimally invasive, without requiring inter-agent communication or cooperative behavior. The key insight behind LiveNet is a unified CBF formulation for simultaneous safety and liveness, which we integrate within a neural network for robustness. We evaluated LiveNet in simulation and found that general multi-robot optimization- and learning-based navigation methods fail to even reach the goal, and while methods designed specially for such environments do succeed, they are 10-20 times slower, 4-5 times more invasive, and much less robust to variations in the scenario configuration such as changes in the start states and goal states, among others. We open-source the LiveNet code at https://github.com/srikarg89/LiveNet{https://github.com/srikarg89/LiveNet.

Graphical User Interfaces (GUIs) are critical to human-computer interaction, yet automating GUI tasks remains challenging due to the complexity and variability of visual environments. Existing approaches often rely on textual representations of GUIs, which introduce limitations in generalization, efficiency, and scalability. In this paper, we introduce Aguvis, a unified pure vision-based framework for autonomous GUI agents that operates across various platforms. Our approach leverages image-based observations, and grounding instructions in natural language to visual elements, and employs a consistent action space to ensure cross-platform generalization. To address the limitations of previous work, we integrate explicit planning and reasoning within the model, enhancing its ability to autonomously navigate and interact with complex digital environments. We construct a large-scale dataset of GUI agent trajectories, incorporating multimodal reasoning and grounding, and employ a two-stage training pipeline that first focuses on general GUI grounding, followed by planning and reasoning. Through comprehensive experiments, we demonstrate that Aguvis surpasses previous state-of-the-art methods in both offline and real-world online scenarios, achieving, to our knowledge, the first fully autonomous pure vision GUI agent capable of performing tasks independently without collaboration with external closed-source models. We open-sourced all datasets, models, and training recipes to facilitate future research at https://aguvis-project.github.io/.

Intelligent autonomous agents hold much potential for the domain of cyber security. However, due to many state-of-the-art approaches relying on uninterpretable black-box models, there is growing demand for methods that offer stakeholders clear and actionable insights into their latent beliefs and motivations. To address this, we evaluate Theory of Mind (ToM) approaches for Autonomous Cyber Operations. Upon learning a robust prior, ToM models can predict an agent's goals, behaviours, and contextual beliefs given only a handful of past behaviour observations. In this paper, we introduce a novel Graph Neural Network (GNN)-based ToM architecture tailored for cyber-defence, Graph-In, Graph-Out (GIGO)-ToM, which can accurately predict both the targets and attack trajectories of adversarial cyber agents over arbitrary computer network topologies. To evaluate the latter, we propose a novel extension of the Wasserstein distance for measuring the similarity of graph-based probability distributions. Whereas the standard Wasserstein distance lacks a fixed reference scale, we introduce a graph-theoretic normalization factor that enables a standardized comparison between networks of different sizes. We furnish this metric, which we term the Network Transport Distance (NTD), with a weighting function that emphasizes predictions according to custom node features, allowing network operators to explore arbitrary strategic considerations. Benchmarked against a Graph-In, Dense-Out (GIDO)-ToM architecture in an abstract cyber-defence environment, our empirical evaluations show that GIGO-ToM can accurately predict the goals and behaviours of various unseen cyber-attacking agents across a range of network topologies, as well as learn embeddings that can effectively characterize their policies.

Multi-Agent Path Finding (MAPF) deals with finding conflict-free paths for a set of agents from an initial configuration to a given target configuration. The Lifelong MAPF (LMAPF) problem is a well-studied online version of MAPF in which an agent receives a new target when it reaches its current target. The common approach for solving LMAPF is to treat it as a sequence of MAPF problems, periodically replanning from the agents' current configurations to their current targets. A significant drawback in this approach is that in MAPF the agents must reach a configuration in which all agents are at their targets simultaneously, which is needlessly restrictive for LMAPF. Techniques have been proposed to indirectly mitigate this drawback. We describe cases where these mitigation techniques fail. As an alternative, we propose to solve LMAPF problems by solving a sequence of modified MAPF problems, in which the objective is for each agent to eventually visit its target, but not necessarily for all agents to do so simultaneously. We refer to this MAPF variant as Transient MAPF (TMAPF) and propose several algorithms for solving it based on existing MAPF algorithms. A limited experimental evaluation identifies some cases where using a TMAPF algorithm instead of a MAPF algorithm with an LMAPF framework can improve the system throughput significantly.

Balancing individual specialisation and shared behaviours is a critical challenge in multi-agent reinforcement learning (MARL). Existing methods typically focus on encouraging diversity or leveraging shared representations. Full parameter sharing (FuPS) improves sample efficiency but struggles to learn diverse behaviours when required, while no parameter sharing (NoPS) enables diversity but is computationally expensive and sample inefficient. To address these challenges, we introduce HyperMARL, a novel approach using hypernetworks to balance efficiency and specialisation. HyperMARL generates agent-specific actor and critic parameters, enabling agents to adaptively exhibit diverse or homogeneous behaviours as needed, without modifying the learning objective or requiring prior knowledge of the optimal diversity. Furthermore, HyperMARL decouples agent-specific and state-based gradients, which empirically correlates with reduced policy gradient variance, potentially offering insights into its ability to capture diverse behaviours. Across MARL benchmarks requiring homogeneous, heterogeneous, or mixed behaviours, HyperMARL consistently matches or outperforms FuPS, NoPS, and diversity-focused methods, achieving NoPS-level diversity with a shared architecture. These results highlight the potential of hypernetworks as a versatile approach to the trade-off between specialisation and shared behaviours in MARL.