simulation

Real-world deployment of intelligent agents is often hindered by scarce training data and the difficulty of constructing high-fidelity simulation environments. To address this, we propose IMAC, the first framework to integrate Unsupervised Environment Design (UED) into world-model-based imagined environments for adaptive, curriculum-driven training from offline data. IMAC synergistically combines world models, UED, and reinforcement learning to conduct progressive, curriculum-style imagination training within procedurally generated latent spaces. Experiments demonstrate that even a lightweight world model trained solely on narrow-domain offline data enables strong zero-shot transfer to unseen environments—validating the feasibility of leveraging compact world models for generalizable agent training. Our core contribution lies in pioneering the joint use of UED and world models for automated curriculum generation, significantly enhancing cross-environment generalization performance.

To address the challenges of difficult development, slow iteration, and high engineering barriers for embodied AI in Minecraft, this paper introduces the first integrated AI agent development platform specifically designed for Minecraft. The platform systematically unifies seven core components: a simulator (Malmo/Obelisk), a high-quality dataset, a model zoo, offline pre-training, online fine-tuning, a lightweight inference engine, and standardized benchmarking suites. Built on a modular Python architecture with a unified API, it natively supports both behavior cloning and reinforcement learning pipelines. Its key innovation lies in bridging the traditional silos among simulation, data, training, and evaluation—enabling a fully reproducible, end-to-end development loop. Experiments demonstrate that the platform substantially reduces agent development time. Accompanied by fully open-sourced code, comprehensive documentation, and step-by-step tutorials, it has already empowered multiple universities and research labs to efficiently validate and advance embodied AI algorithms.

To address the inefficiency of simulation-based training and the difficulty of sim-to-real transfer in robot learning, this paper introduces the first fully open-source, integrated framework built upon MJX (JAX-accelerated MuJoCo), unifying physics simulation, batched GPU rendering, and reinforcement learning training. The framework supports diverse robotic morphologies—including quadrupeds, humanoid robots, dexterous hands, and manipulators—and enables end-to-end policy training in minutes on a single GPU. Crucially, it achieves cross-modal zero-shot sim-to-real transfer—demonstrated for both state-based and pixel-based inputs—within a unified open platform; pre-trained policies deploy directly onto real-world quadrupeds and robotic arms without fine-tuning. All code, pretrained models, and demonstration videos are publicly released at playground.mujoco.org.

Non-AI engineers lack efficient tools to prototype generative AI-based UI Agent experiences. Method: We propose AgentBuilder—a user-centered design framework and lightweight prototyping tool—developed through contextual inquiry, design probes, and in-situ experiments to identify core activities and capability requirements for agent experience design. AgentBuilder supports low-code interaction orchestration, real-time LLM integration, and iterative multi-turn dialogue prototyping. Evaluation with 14 cross-disciplinary participants demonstrated that the framework significantly lowers prototyping barriers, enhances designer engagement, and improves feedback quality. It further revealed non-technical users’ critical needs for controllability, explainability, and progressive guidance. Contribution/Results: This work presents the first systematic methodology and practical toolchain for UI Agent experience prototyping tailored specifically for non-engineers, bridging a critical gap between human-centered design and generative AI interface development.

Traditional manual 3D modeling is inefficient and insufficient for constructing large-scale, dynamic, interactive 3D worlds. Method: This paper proposes a lightweight multimodal framework that tightly integrates LLaMA-2-7B with Unreal Engine 5’s rendering pipeline, enabling end-to-end 3D scene generation from text or visual instructions. A lightweight LLM parses multimodal inputs and coordinates with a physics engine to achieve high-fidelity dynamic simulation, while UE5 delivers real-time rendering and agent interaction. Contribution/Results: We introduce the first scalable, interactive, and dynamically evolving multimodal 3D generation paradigm. Our approach significantly outperforms baselines in layout accuracy and visual fidelity, improves production efficiency by over 90×, and maintains high creative controllability. The framework is applicable to critical domains including embodied AI and autonomous driving simulation.

Existing static evaluation methods struggle to effectively assess the agentic capabilities of large language models in multi-step decision-making tasks. To address this limitation, this work proposes AgenticAI-Supervisor, the first reinforcement learning simulation framework for agentic AI that supports closed-loop feedback. By decoupling environment construction from scalable execution, the framework integrates API- and UI-driven Gym environments, generates high-fidelity execution trajectories, employs multi-dimensional reward shaping, and incorporates internal state validation mechanisms to mitigate reward hacking. Demonstrated in a customer service agent case study, the framework enables stable closed-loop feedback and significantly enhances model optimization outcomes.

Current agent research is constrained by simplistic software environments and short-horizon tasks, lacking scalable methods for constructing complex, realistic settings. This work proposes Gym-Anything, a framework that automatically transforms arbitrary software into interactive agent environments. It innovatively formulates environment construction as a multi-agent collaborative task, integrating code generation, vision-language models (VLMs), real-world data configuration, trajectory distillation, and an automated auditing mechanism. Leveraging the GDP occupational classification, we introduce CUA-World—a benchmark comprising over 10,000 long-horizon tasks, with CUA-World-Long episodes exceeding 500 steps. The distilled 2B-parameter VLM outperforms models twice its size, and the auditing feedback loop improves Gemini-3-Flash’s success rate from 11.5% to 14.0%.

This work addresses the scarcity of diverse, automatically constructible interactive 3D training environments for embodied agents by introducing the first open-source platform that enables co-evolution between environment generation and embodied learning. Built on Unreal Engine 5, the platform integrates large language and vision models, engine-level scripting, Gym-compatible interfaces, physics validation, and vision-language model (VLM) feedback. It features SimCoder, a tool- and skill-augmented coding agent that autonomously generates physically plausible, executable 3D environments from textual or visual instructions and dynamically presents challenging tasks near the agent’s current capability boundary. Experiments demonstrate that this self-evolving framework substantially improves environment generation reliability, boosting agent success rates on unseen navigation benchmarks by 18% over fixed environments and by 40% compared to untrained baselines.

This study addresses the challenge that current AI coding agents in scientific software development often produce outputs that pass conventional tests yet violate physical principles, making it difficult to distinguish symptom mitigation from genuine root-cause resolution. Under the supervision of physicists, the authors leveraged Claude-series models and the JAX framework to develop CLAX-PT—a differentiable first-order perturbation theory module—within 12 days, documenting 15 critical human interventions. Through oracle testing, multi-parameter validation, shared logs, and physics-informed constraints, they observed that the AI repeatedly optimized flawed architectures across 33 sessions until explicit injection of physical concepts triggered effective refactoring. The findings underscore that thoughtfully designed supervision mechanisms are more critical than raw model capability and identify three actionable practices to bridge testing blind spots and ensure the reliability of scientific computation.

This study addresses the pervasive overestimation of Computer-Using Agent (CUA) performance in existing evaluations, which stems from poorly designed environments and statistically unreliable methodologies. To rectify this, the work systematically uncovers the issue for the first time and introduces PRISM—a set of five principled guidelines for robust environment design. It further presents DigiWorld, a sandboxed benchmark comprising 15 mobile applications and over 3.2 million verified configurations. Complementing this, the authors propose a statistically sound evaluation framework based on Wilson confidence intervals and hierarchical bootstrap resampling to enable trustworthy assessment of CUA capabilities. Empirical results demonstrate that conventional static benchmarks are easily outperformed by trivial replay scripts, thereby underscoring the necessity and efficacy of the proposed evaluation paradigm.