X-Policy-Lab：面向机器人策略评估与部署的统一标准与开放生态系统【集成 42 个机器人策略】

XPolicyLab: A Unified Standard and Open Ecosystem for Robot Policy Evaluation and DeploymentXPolicyLab Contributors, Leading Institutions: MMLab@HKU THU Website: https://xpolicylab.github.io/XPolicyLabStandard and Infra for Policy Evaluation and Deployment XPolicyLab.github.ioFig. 1: Overview of XPolicyLab. A standard and infrastructure for robot policy evaluation and deployment.Abstract— Robot policy evaluation and deployment remain fragmented by model-specific software dependencies, data representations, and runtime interfaces, so that connecting N policies to M evaluation environments requires O(NM) separate integrations. We present XPolicyLab, a unified standard and open ecosystem that reduces this cost to O(N+M). XPolicyLab specifies common observation, action, and trajectory schemas together with a minimal adapter interface for observation updates, action prediction, batched execution, and episode reset, while a dependency-isolated client/server architecture separates policy inference from environment execution, so that each side retains its native software stack and may run locally or remotely. The ecosystem integrates 42 robot policies and standardizes their installation, debugging, serving, and evaluation workflows. Across these adapters, model-specific code varies by an order of magnitude while the environment-facing loop stays within a few lines of a fixed reference, confirming that the contract confines heterogeneity to the policy side. In a controlled study, conforming to the standard reduces the integration effort of a representative policy from over five hours to two hours, and packaged agent skills reduce it further to thirty minutes. The same adapters serve RoboTwin, RoboDojo simulation, and standardized realrobot evaluation through one interface. XPolicyLab is released as shared infrastructure for reproducible policy comparison and standardized deployment across simulation and physical platforms. Project website: https://xpolicylab.github.io/.XPolicyLab：面向机器人策略评估与部署的统一标准与开放生态系统XPolicyLab 贡献者牵头机构：MMLab@HKU THU网站：https://xpolicylab.github.io/面向 **策略评估与部署（Policy Evaluation and Deployment）** 的标准与基础设施图 1：XPolicyLab 概览。面向 **机器人策略评估与部署（Robot Policy Evaluation and Deployment）** 的标准与基础设施。摘要（Abstract）——机器人策略评估与部署（Robot Policy Evaluation and Deployment）目前仍因模型特定的软件依赖、数据表示方式以及运行时接口而呈碎片化状态，因此，将N NN个策略连接到M MM个评估环境需要进行O ( N M ) O(NM)O(NM)次独立集成。我们提出 XPolicyLab，这是一个统一标准与开放生态系统（Unified Standard and Open Ecosystem），可将这一成本降低至O ( N + M ) O(N+M)O(N+M)。XPolicyLab 规定了通用的观测、动作与轨迹模式（Observation, Action, and Trajectory Schemas），并定义了一套最小化的适配器接口（Adapter Interface），用于观测更新、动作预测、批量执行以及回合重置。同时，一种依赖隔离的客户端/服务器架构（Dependency-Isolated Client/Server Architecture）将策略推理（Policy Inference）与环境执行（Environment Execution）分离，使两侧均能够保留各自原生的软件栈，并可在本地或远程运行。该生态系统集成了42 个机器人策略，并对这些策略的安装、调试、服务化以及评估工作流程进行了标准化。在这些适配器中，模型特定代码的规模相差可达一个数量级，而面向环境的执行循环始终保持在固定参考实现的几行代码以内，这证明该接口契约能够将异质性限制在策略侧。在一项受控研究中，遵循该标准后，一个具有代表性的策略的集成工作量从超过五小时减少至两小时；进一步使用封装好的智能体技能（Agent Skills）后，所需时间可继续降低至三十分钟。同一套适配器可通过统一接口支持 RoboTwin、RoboDojo 仿真以及标准化真实机器人评估（Standardized Real-Robot Evaluation）。XPolicyLab 作为共享基础设施发布，旨在支持跨仿真平台与物理平台的可复现策略比较（Reproducible Policy Comparison）与标准化部署（Standardized Deployment）。项目网站：https://xpolicylab.github.io/I. INTRODUCTIONGeneralist robot policies, including vision-language-action (VLA) models and other embodied foundation models, have rapidly expanded in capability and task coverage [1], [2], [3], [4], [5], [6], [7], [8]. Their evaluation and deployment, however, remain constrained by fragmented system stacks [9]. Different policies assume incompatible software dependencies, observation and action representations, checkpoint conventions, and runtime interfaces, so deploying a policy on a new benchmark, simulator, or physical robot requires substantial platform-specific engineering. Observation preprocessing, action conversion, control loops, and runtime orchestration are repeatedly reimplemented, which increases integration effort and, more damagingly, silently re-fixes conventions such as camera naming, channel order, and gripper scaling, so that reported results become difficult to compare across policies. Stated as a systems problem, connecting N policies to M evaluation environments currently costs O(NM) separate integrations.Substantial progress has been made in robot data collection, dataset standardization, and training infrastructure [10], [11], [12], [13], [14], which lowers the cost of producing a checkpoint. The complementary boundary, between a trained policy and the environments that must execute it, remains largely unaddressed. A reusable solution there must absorb heterogeneous model architectures and dependencies while exposing a stable interface to simulators, benchmark clients, and physical robots. In particular, it must support stateful policies, action chunking, batched evaluation, local and remote execution, and reliable communication, without coupling policy implementations to environment-specific software stacks.We present XPolicyLab,1a unified standard and open ecosystem for robot policy evaluation and deployment, illustrated in Fig. 2. XPolicyLab specifies common observation, action, and trajectory representations, together with a lightweight adapter contract for policy initialization, observation updates, action prediction, batched execution, and episode reset, so that each integrated policy retains its native implementation, dependencies, checkpoints, and training procedure. A dependency-isolated serving architecture separates policy inference from environment execution through a common communication layer, allowing policy servers and environment clients to run in independent software environments, either on the same machine or across a network. Around this core, XPolicyLab standardizes the workflows for installation, data conversion, training, serving, debugging, and evaluation, and packages the conformance procedure as machine-readable agent skills that a coding agent executes to scaffold, implement, and validate a new adapter.As of August 2026, XPolicyLab integrates 42 robot policies, spanning VLA models, world-action models, diffusionbased visuomotor policies, memory-augmented policies, and conventional imitation-learning baselines. By separating model-facing adapters from environment-facing clients, a single adapter is reused across benchmark ecosystems, simulation backends, and physical robot evaluation systems without modifying the underlying policy implementation. We demonstrate this portability through deployments on RoboTwin [15] and RoboDojo [16], including standardized physical evaluation through RoboDojo-RealEval, for which XPolicyLab also provides the policy submission and execution workflow of the official public leaderboards.Our main contributions are summarized as follows:We introduce a unified contract between robot policies and evaluation environments that standardizes observation and action representations, action prediction, batched execution, and episode management while preserving policy-specific implementations.We develop a dependency-isolated serving architecture that decouples policy inference from environment execution and supports local and remote deployment across simulators, benchmark clients, and physical robot systems.We package the conformance procedure as machinereadable agent skills, so that scaffolding, implementing, and auditing a new adapter against the standard becomes a guided and checkable procedure rather than a documentation-reading exercise.We establish an open ecosystem comprising 42 robot policies across diverse policy families, and evaluate it in terms of policy coverage, the locality of integration complexity, integration effort, and adapter reuse, with deployments on RoboTwin [15] and RoboDojo [16].I. 引言通用型机器人策略（Generalist Robot Policies），包括视觉-语言-动作模型（Vision-Language-Action Models, VLA）以及其他具身基础模型（Embodied Foundation Models），其能力和任务覆盖范围近年来迅速扩展 [1], [2], [3], [4], [5], [6], [7], [8]。然而，这些模型的评估与部署（Evaluation and Deployment）仍受到碎片化系统栈（System Stacks）的制约 [9]。不同策略依赖彼此不兼容的软件依赖、观测与动作表示、检查点约定以及运行时接口，因此，将一个策略部署到新的基准、仿真器或物理机器人上，需要大量针对具体平台的工程工作。观测预处理（Observation Preprocessing）、动作转换（Action Conversion）、控制循环（Control Loops）和运行时编排（Runtime Orchestration）被反复重新实现，这不仅增加了集成工作量，更严重的是，还会在不易察觉的情况下重新固定诸如相机命名、通道顺序以及夹爪缩放等约定，从而使不同策略之间报告的结果难以进行比较。从系统问题的角度来看，目前将N NN个策略连接到M MM个评估环境需要进行O ( N M ) O(NM)O(NM)次独立集成。在机器人数据采集（Robot Data Collection）、数据集标准化（Dataset Standardization）和训练基础设施（Training Infrastructure）方面，已经取得了显著进展 [10], [11], [12], [13], [14]，从而降低了生成一个检查点（Checkpoint）的成本。然而，与之互补的另一边界——即经过训练的策略与必须执行该策略的环境之间的边界——在很大程度上仍未得到解决。针对这一边界的可复用解决方案必须能够容纳异构的模型架构和依赖，同时向仿真器、基准客户端以及物理机器人暴露稳定的接口。尤其是，它必须支持有状态策略（Stateful Policies）、动作分块（Action Chunking）、批量评估（Batched Evaluation）、本地与远程执行以及可靠通信，同时又不能将策略实现与特定环境的软件栈耦合在一起。我们提出 XPolicyLab1，这是一个用于机器人策略评估与部署（Robot Policy Evaluation and Deployment）的统一标准与开放生态系统（Unified Standard and Open Ecosystem），如图 2 所示。XPolicyLab规定了通用的观测、动作与轨迹表示（Observation, Action, and Trajectory Representations），并定义了一套轻量级的适配器契约（Adapter Contract），用于策略初始化、观测更新、动作预测、批量执行以及回合重置，从而使每个已集成策略都能够保留其原生实现、依赖、检查点和训练流程。一种依赖隔离的服务架构（Dependency-Isolated Serving Architecture）通过统一的通信层，将策略推理（Policy Inference）与环境执行（Environment Execution）分离，使策略服务器与环境客户端能够运行在彼此独立的软件环境中，既可以部署在同一台机器上，也可以跨网络运行。围绕这一核心，XPolicyLab 对安装、数据转换、训练、服务化、调试和评估等工作流程进行了标准化，并将标准符合性流程（Conformance Procedure）封装为机器可读的智能体技能（Machine-Readable Agent Skills），由编程智能体（Coding Agent）执行，以搭建、实现并验证新的适配器。截至 2026 年 8 月，XPolicyLab 已集成 42 个机器人策略，涵盖视觉-语言-动作模型（VLA Models）、世界-动作模型（World-Action Models）、基于扩散的视觉运动策略（Diffusion-Based Visuomotor Policies）、记忆增强策略（Memory-Augmented Policies）以及传统的模仿学习基线（Imitation-Learning Baselines）。通过将面向模型的适配器（Model-Facing Adapters）与面向环境的客户端（Environment-Facing Clients）分离，同一个适配器可以在不同的基准生态系统、仿真后端以及物理机器人评估系统之间复用，而无需修改底层策略实现。我们通过在 RoboTwin [15] 和 RoboDojo [16] 上的部署展示了这种可移植性（Portability），其中包括通过 RoboDojo-RealEval 进行的标准化物理机器人评估（Standardized Physical Evaluation）。对于 RoboDojo-RealEval，XPolicyLab 还提供了官方公开排行榜所采用的策略提交与执行工作流程（Policy Submission and Execution Workflow）。我们的主要贡献总结如下：我们提出了一套连接机器人策略与评估环境的统一契约（Unified Contract），对观测与动作表示、动作预测、批量执行以及回合管理进行标准化，同时保留各策略自身特定的实现。我们开发了一种依赖隔离的服务架构（Dependency-Isolated Serving Architecture），将策略推理与环境执行解耦，并支持在仿真器、基准客户端以及物理机器人系统之间进行本地和远程部署。我们将标准符合性流程（Conformance Procedure）封装为机器可读的智能体技能（Machine-Readable Agent Skills），使得针对该标准搭建、实现和审计一个新的适配器，成为一个有引导且可检查的流程，而不再只是一个阅读文档的过程。我们建立了一个开放生态系统，其中包含来自不同策略类别的 42 个机器人策略，并从策略覆盖范围（Policy Coverage）、集成复杂度的局部性（Locality of Integration Complexity）、集成工作量（Integration Effort）以及适配器复用（Adapter Reuse）等方面对其进行了评估，同时在 RoboTwin [15] 和 RoboDojo [16] 上完成了部署。II. RELATED WORKA. Robot Data and Training InfrastructureLarge-scale datasets and open-source frameworks have improved the availability of robot demonstrations and the reproducibility of policy training. DROID and BridgeData provide diverse real-world manipulation data [12], [17], FastUMI supports embodiment-agnostic data collection [11], LeRobot standardizes dataset processing and policy training [10], and RLinf provides scalable infrastructure for policy optimization and execution [14]. These systems target data acquisition, dataset organization, and model optimization, and therefore lower the cost of producing a checkpoint. XPolicyLab addresses the complementary boundary of how an already-trained policy is integrated, served, evaluated, and deployed across heterogeneous environment stacks.B. Robot Policy Evaluation PlatformsSimulation benchmarks such as RLBench, ManiSkill2, LIBERO, RMBench, and UniVTAC provide controlled environments for evaluating robotic skills and generalization [18], [19], [20], [21], [22]. RoboTwin extends this setting to diverse bimanual manipulation tasks [23], [15], [24], while RoboDojo combines capability-oriented simulation with standardized physical evaluation [16]. Beyond simulation, RoboArena and RoboChallenge coordinate distributed real-robot evaluation across institutions and hardware [25], [26]. These platforms define task semantics, environment interfaces, evaluation protocols, and performance metrics; that is, they specify what is evaluated and under which conditions. XPolicyLab does not replace them, and instead supplies the policy-side abstraction that lets heterogeneous policies reach any of them through one interface, so that no benchmark needs to maintain its own model-specific serving stack.C. Robot Policy DeploymentRobot policies such as OpenVLA,π 0 , \pi _ { 0 } ,π0​,and GR00T ship inference and deployment code tailored to their respective architectures, dependencies, observation formats, and action representations [1], [27], [28]. Each such stack is faithful to one model, which is precisely why deploying a policy across multiple benchmarks, simulators, and physical robots requires repeated adaptation of preprocessing pipelines, action conversion, runtime environments, and control loops, and why the community pays the O(NM) cost identified above. XPolicyLab removes this pairwise dependency through standardized observation and action schemas, a minimal adapter interface, and dependency-isolated serving. Unlike a unified modeling framework, it prescribes no network architecture, training objective, action decoder, or temporal horizon, which is what allows adapters to be reused across evaluation environments while model-specific implementations remain intact.以下为严格中文翻译：II. 相关工作A. 机器人数据与训练基础设施大规模数据集与开源框架提升了机器人示范数据的可获得性，并改善了策略训练的可复现性。DROID 和 BridgeData 提供了多样化的真实世界操作数据 [12], [17]；FastUMI 支持与具身形态无关的数据采集（Embodiment-Agnostic Data Collection）[11]；LeRobot 对数据集处理（Dataset Processing）与策略训练（Policy Training）进行了标准化 [10]；RLinf 则提供了用于策略优化与执行（Policy Optimization and Execution）的可扩展基础设施 [14]。这些系统主要面向数据获取、数据集组织和模型优化，因此降低了生成一个检查点（Checkpoint）的成本。XPolicyLab所解决的是与之互补的另一侧边界，即一个已经训练完成的策略如何跨异构环境栈进行集成、服务化、评估与部署。B. 机器人策略评估平台RLBench、ManiSkill2、LIBERO、RMBench 和 UniVTAC 等仿真基准（Simulation Benchmarks）提供了受控环境，用于评估机器人技能及其泛化能力 [18], [19], [20], [21], [22]。RoboTwin 将这一评估设定扩展至多样化的双臂操作任务（Bimanual Manipulation Tasks）[23], [15], [24]，RoboDojo 则将面向能力的仿真（Capability-Oriented Simulation）与标准化物理评估（Standardized Physical Evaluation）相结合 [16]。在仿真之外，RoboArena 和 RoboChallenge 面向不同机构和硬件平台，协调开展分布式真实机器人评估（Distributed Real-Robot Evaluation）[25], [26]。这些平台定义了任务语义、环境接口、评估协议与性能指标；也就是说，它们规定了“评估什么”以及“在何种条件下进行评估”。XPolicyLab 并不取代这些平台，而是提供策略侧抽象（Policy-Side Abstraction），使异构策略能够通过统一接口接入任意此类平台，从而使各个基准无需维护各自特定于模型的服务栈。C. 机器人策略部署OpenVLA、π 0 \pi_{0}π0​和 GR00T 等机器人策略通常附带针对各自架构、依赖、观测格式以及动作表示而定制的推理与部署代码 [1], [27], [28]。每一套此类软件栈都忠实服务于单一模型，而这恰恰意味着，当一个策略需要跨多个基准、仿真器和物理机器人进行部署时，就必须反复适配预处理流水线（Preprocessing Pipelines）、动作转换（Action Conversion）、运行时环境（Runtime Environments）以及控制循环（Control Loops），这也是为什么整个社区需要承担前文所指出的O ( N M ) O(NM)O(NM)集成成本。XPolicyLab 通过标准化的观测与动作模式、最小化适配器接口以及依赖隔离的服务机制（Dependency-Isolated Serving），消除了这种成对依赖关系。与统一建模框架（Unified Modeling Framework）不同，XPolicyLab 并不规定任何网络架构、训练目标、动作解码器或时间范围；正是这种设计，使得适配器能够跨不同评估环境进行复用，同时仍完整保留各模型特定的实现。III. XPOLICYLABXPolicyLab adopts a policy-centric abstraction that decouples policy computation from environment execution. Its central principle is that each policy exposes a stable external contract while retaining control over its architecture, dependencies, checkpoints, and inference procedure, so that heterogeneous benchmarks, simulators, and physical robots can invoke any policy through a common execution pathway.XPolicyLab 采用以策略为中心的抽象（Policy-Centric Abstraction），将策略计算（Policy Computation）与环境执行（Environment Execution）解耦。其核心原则是：每个策略在保留对自身架构、依赖、检查点和推理流程控制权的同时，对外暴露稳定的外部契约（External Contract），从而使异构的基准、仿真器和物理机器人能够通过统一的执行路径调用任意策略。Fig. 2: Infrastructure of XPolicyLab. Heterogeneous policy runtimes (left) keep their native dependencies, checkpoints, and processing pipelines, while evaluation environments (right) keep their simulator and robot stacks. XPolicyLab (center) joins the two through a unified lifecycle from installation to evaluation, a minimal adapter contract, standardized observation and action schemas, and a dependency-isolated serving bridge. A single adapter therefore serves public benchmarks, simulation backends, and physical robots, under local or remote execution and batched, stateful inference.图 2：XPolicyLab 的基础设施。异构的策略运行时（Policy Runtimes）（左侧）保留各自原生的依赖、检查点和处理流水线，而评估环境（Evaluation Environments）（右侧）则保留各自的仿真器与机器人软件栈。XPolicyLab（中间）通过从安装到评估的统一生命周期、最小化适配器契约、标准化的观测与动作模式，以及依赖隔离的服务桥接机制（Dependency-Isolated Serving Bridge）将二者连接起来。因此，单个适配器即可服务于公开基准、仿真后端和物理机器人，并支持本地或远程执行，以及批量、有状态的推理。A. System OverviewXPolicyLab comprises four functional components: an environment backend that executes tasks, an environment client that manages the interaction loop, a communication layer that transfers observations and actions, and a policy server that hosts the model and its adapter. Cross-cutting workflow tooling standardizes installation, configuration, debugging, training, serving, and evaluation. The policy server executes in an independent software environment and may reside either on the same machine as the environment client or on a remote compute node.The architecture establishes two stable boundaries. A semantic boundary specifies the shared observation and action schemas, whereas an execution boundary defines communication between policy servers and environment clients. For an observationo t . \mathbf { o } _ { t } .ot​., an optional internal states t , \mathbf { s } _ { t } ,st​,and an action horizon H, policy inference can be expressed asa t : t + H − 1 = g o u t ( π θ ( g i n ( o t ) , s t ) ) , (1) \mathbf { a } _ { t : t + H - 1 } = g _ { \mathrm { o u t } } \left( \pi _ { \theta } \left( g _ { \mathrm { i n } } ( \mathbf { o } _ { t } ) , \mathbf { s } _ { t } \right) \right) ,\tag{1}a