具身智能开源合辑：从数据集到仿真平台覆盖全链路

2026年07月29日 11:41

本文共计18135个字，预计阅读时长61分钟。

来源/具身研习社责编/NuanxinXiaoMo 暖心小茉

前言

过去半年，具身智能领域的开源节奏明显加快。对数据集进行了深入分析，仿真器与模型框架以及硬件平台得到了开发与优化。来自“数据燃料”到“机器人大脑”再到“物理身体”的开源成果几乎覆盖了全链路。然而，如果要为这半年的开源浪潮找一个最集中的爆发点，那一定是 ICRA 2026——这个全球机器人学界与工业界的顶级会议，几乎成了各家开源动作的首发场。

In contrast to the surge in paper publications, what merits closer scrutiny is the wave of open-source projects densely released in their wake. These efforts span the entire closed-loop process, encompassing data acquisition, simulation verification, and real-world deployment. This article, grounded in this open-source momentum, eschews exhaustive indexing. Instead, it selects one or two flagship projects per category for in-depth examination of their value, while supplying a comprehensive resource table for each category to guide readers to the relevant resources.

Figure 1: A scene with high-degree-of-freedom dual-arm hands in real-world data acquisition, covering tasks such as grasping, assembly, and manipulation of articulated objects.

关键点速览

开源数据集正在定义具身智能的燃料标准。

仿真平台——让Sim2Real从“鸿沟”变“坦途”

模型与框架——VLA从“黑箱”走向“思考”

In hardware and algorithm tools, the final mile of implementation has been realized.

Effects verification—when 526 teams share one set of open-source toolchains.

True-world data acquisition incurs prohibitively high costs, which constrains generalization in embodied intelligence. Consequently, whoever releases high-quality data thereby defines the foundational infrastructure of the field.

AGIBOT WORLD dataset from Zhiyuan Robotics undoubtedly represents the most heavyweight open-source initiative at this conference. This dataset incorporates over one million high-quality trajectories obtained through both real-world acquisition and simulation generation. It encompasses complex scenarios such as logistics sorting, workpiece flipping, shelf organization, popcorn scooping, door opening, desk cleaning, and dual-arm pot operation. Each task includes hundreds of complete operational trajectories.

This is not only the most advanced real-world data accumulation in the industry at present, but also the starting point for Zhiyuan's data-simulation-real-world closed loop. The dataset has been fully open-sourced in the two communities of Hugging Face and ModelScope.

另一个值得关注的数据集是 MERL（三菱电机研究所）的UNIC多模态接触估计数据集。

In the industry, contact estimation (namely the determination of the interaction state between the grasped object and the external environment) remains one of the hardest obstacles that VLA models have yet to overcome. The UNIC framework encodes visual observations and integrates proprioceptive and tactile modalities without any prior knowledge or camera calibration. The accompanying dataset amounts to approximately 98 GB and is released under the CC-BY-SA-4.0 license. The code repository and data are both open-sourced.

Besides the two representative projects mentioned above, the following dataset is also worthy of attention:

获取方式：

AGIBOT WORLD数据集地址：

AGIBOT WORLD Challenge 2026 dataset from Zhiyuan Robotics unquestionably represents the most heavyweight open-source initiative at this conference.
This dataset incorporates over one million high-quality trajectories obtained through both real-world acquisition and simulation generation.
It encompasses complex scenarios such as logistics sorting, workpiece flipping, shelf organization, popcorn scooping, door opening, desk cleaning, and dual-arm pot operation.
Each task includes hundreds of complete operational trajectories.
This dataset has been fully open-sourced in the two communities of Hugging Face and ModelScope.

AgiBotWorldChallenge-2026 dataset from Zhiyuan Robotics unquestionably represents the most heavyweight open-source initiative at this conference. This dataset incorporates over one million high-quality trajectories obtained through both real-world acquisition and simulation generation. It encompasses complex scenarios such as logistics sorting, workpiece flipping, shelf organization, popcorn scooping, door opening, desk cleaning, and dual-arm pot operation. Each task includes hundreds of complete operational trajectories. This dataset has been fully open-sourced in the two communities of Hugging Face and ModelScope.

UNIC数据集地址：

The AGIBOT WORLD dataset from Zhiyuan Robotics unquestionably represents the most heavyweight open-source initiative at this conference.
This dataset incorporates over one million high-quality trajectories obtained through both real-world acquisition and simulation generation.
It encompasses complex scenarios such as logistics sorting, workpiece flipping, shelf organization, popcorn scooping, door opening, desk cleaning, and dual-arm pot operation.
Each task includes hundreds of complete operational trajectories.
This dataset has been fully open-sourced in the two communities of Hugging Face and ModelScope.

The AGIBOT WORLD dataset from Zhiyuan Robotics undoubtedly represents the most heavyweight open-source initiative at this conference. This dataset incorporates over one million high-quality trajectories obtained through both real-world acquisition and simulation generation. It encompasses complex scenarios such as logistics sorting, workpiece flipping, shelf organization, popcorn scooping, door opening, desk cleaning, and dual-arm pot operation. Each task includes hundreds of complete operational trajectories.

Another dataset worthy of attention is the UNIC multimodal contact estimation dataset from MERL.

In the industry, contact estimation remains one of the hardest obstacles that VLA models have yet to overcome. The UNIC framework encodes visual observations and integrates proprioceptive and tactile modalities without any prior knowledge or camera calibration. The accompanying dataset amounts to approximately 98 GB and is released under the CC-BY-SA-4.0 license. The code repository and data are both open-sourced.

Besides the two representative projects mentioned above, the following dataset is also worthy of attention:

The UNIC dataset Part 2 from MERL. This dataset is hosted on Zenodo at https://zenodo.org/records/20287722 and provides the multimodal sensory data used in the UNIC: Learning Unified Multimodal Extrinsic Contact Estimation work. The training dataset is a Zarr archive of approximately 98 GB. For distribution it is split into two balanced zip parts, each about 48.9 GB.

WideDepth数据集地址： https://ilyaind.github.io/WideDepth

EgoTraj-Bench数据集地址： https://github.com/zoeyliu1999/EgoTraj-Bench

野外深度补全数据集地址： https://github.com/ntnu-frl/depth-completion

ArtVIP数据集地址： https://huggingface.co/datasets/X-Humanoid/ArtVIP

仿真平台——让Sim2Real从“鸿沟”变“坦途”

Sim2Real translation remains painful due to insufficiently realistic simulations. The two simulation platforms at ICRA 2026 approached the same goal from different paths: making Sim2Real translation approach zero cost.

智元Genie Sim 3.0 is the first large language model driven open-source simulation platform in the industry. It integrates 3D reconstruction and vision generation technologies to build high-fidelity digital twin environments. The scenes assets and tasks are consistent with the competition test server allowing participants to complete closed-loop evaluation locally. Its core value is that when large models begin to drive simulation environments scene generation no longer relies on manual construction but on semantic understanding.

中兴通讯RealMirror则走了另一条路。它是一个集“数据采集-模型训练-模型推理-自动化评测-零样本Sim2Real迁移”于一体的端到端VLA仿真基座。该平台覆盖了5大场景，包括厨房清理、空气炸锅操作、流水线分拣以及杯间倾倒等，同时具备5大核心技能，比如双臂协同以及动态抓取等。这些功能累计提供了1200条高质量仿真轨迹。

In terms of technical metrics, RealMirror achieves over 95 percent visual rendering accuracy compared to real scenes and maintains physical parameter errors within 5 percent. After its open-source release, it quickly surpassed 600 GitHub stars, with over 2400 downloads of its 120 GB open-source assets. Whereas Genie Sim makes simulation more real, RealMirror achieves zero-cost Sim2Real translation, enabling models trained solely on simulation data without any real-world fine-tuning to complete tasks directly on physical robots.

Figure 3: RealMirror encompasses scene construction, teleoperation data acquisition, VLA training evaluation, and zero-cost Sim2Real migration.

以下为仿真平台完整清单：

Genie Sim 3.0平台地址：https://agibot-world.com/genie-sim

RealMirror平台地址：https://github.com/terminators2025/RealMirror

模型与框架——VLA从“黑箱”走向“思考”

ICRA 2026 witnessed the evolution of VLA models from black-box outputs to thinking and reasoning, as well as the story of lightweight models challenging the giants in a reversal.

AcoT-VLA, jointly developed by Zhiyuan and Beihang University, serves as one of the baseline models at this conference. Its core innovation lies in introducing a chain of thought, which enables VLA models to possess an explicit process of reasoning and action. Traditional VLA models behave like intuitive players that directly output actions upon seeing a scene. ACoT, in contrast, operates like an analytical player that first thinks about where to grasp and then where to grasp next before executing.

Figure 4: ACoT-VLA constrains action prediction through the joint action of explicit action intent and implicit visual inference.

原力灵机（Dexmal）则是本届大会开源领域最令人惊喜的“黑马”。其自研的 DM0 大模型是全球首个从底层数据到训练范式再到模型架构全部面向真实物理世界机器人作业而设计的具身原生大模型。最令人震撼的是它的“小身材大能量”——仅 2.4B 参数，却在 RoboChallenge 真机评测中位居全球第一。

DM0 incorporates multi-perception data, intelligent driving data, and internet data through joint training with Step Star and Qianli Technology at the data level. It covers three core tasks—dexterous manipulation, environmental navigation, and whole-body control—at the architectural level through joint training across eight different robot bodies. At the capability level, it extends chain-of-thought reasoning into the physical space and achieves operation precision at the sub-millimeter level.

原力灵机同步亮相的 Dexbotic 开源框架则是全球第二、国内唯一的通用具身智能开源框架。它连接数据、模型、训练、评测、硬件与真机部署，让开发者能够以统一流程构建真实可用的机器人智能。自发布以来，该框架已服务清华、北大、普林斯顿、帝国理工等数十家顶尖院校以及腾讯等企业，汇聚超千位开发者共建生态。

其核心特性包括多源数据混合训练、统一操作与导航、统一模仿学习与强化学习，以及模块化解耦架构，这些特性让“训一个能用的具身大模型”从少数团队的手艺活变成标准化可复现的工程问题。

Figure 5: Dexbotic unifies multiple robot embodiments, data processing, model modules, and experimental workflows into a single toolbox.

此外，Dexora作为首个原生面向双臂双高自由度（36自由度）灵巧操作的开源VLA模型，也同期开源。它搭载两台6自由度机械臂与一对12自由度灵巧手，数据集包含10万条仿真轨迹和1万条真实遥操作轨迹。如果说DM0追求通用，Dexora则死磕极致灵巧。

以下为模型与框架完整清单（含经典规划库）。

ACoT-VLA模型地址：https://github.com/AgibotTech/ACoT-VLA

DM0大模型地址：https://github.com/dexmal/dexbotic/blob/main/docs/DM0.md

Dexbotic框架地址：

DM0 represents the flagship open-source VLA model released at this conference. It is built upon a dual-expert architecture that merges attention with Flow Matching to generate continuous robot actions. Unlike conventional CogACT or OFT models, DM0 produces entire action trajectories through a diffusion-based method, yielding chunks of future actions in a single forward pass. The model is supported by pretrained checkpoints that users can download directly from Hugging Face.

Training proceeds in three stages. The first stage conducts large-scale unified pretraining on the vision-language model by integrating web text, autonomous driving scenarios, and embodied interaction logs. This process enables the model to acquire semantic knowledge while simultaneously establishing physical priors such as spatial relationships and dynamics. In the second stage, a Flow Matching-based action expert is constructed atop the vision-language model. To reconcile high-level reasoning with precise low-level control, a hybrid gradient strategy is employed: action-expert gradients are isolated from the vision-language model on embodied data, while the vision-language model continues to learn from non-embodied corpora.

The third stage applies post-training with embodied spatial scaffolding. This mechanism constructs spatial chain-of-thought reasoning that links perception, task understanding, motion planning, and execution into a closed loop. Consequently, the solution space for actions is effectively constrained, enabling sub-millimeter operation precision. The model supports two inference modes. In the first, continuous actions are generated directly from multimodal observations and instructions. In the second, textual reasoning is produced first before the action expert conditions on the generated text.

Hybrid DM0 co-training is also available. It mixes vision-language-action action learning with general vision-language modeling data, allowing the action expert to retain its specialized capabilities while the vision-language model preserves dialogue supervision. The effective vision-language-action to vision-language modeling sampling ratio is controlled by dataset frequencies, and action normalization statistics are computed exclusively from action datasets to prevent interference from non-embodied samples.

At evaluation, DM0 achieves a Libero benchmark average success rate of 94.1 percent across spatial, object, goal, and long-horizon tasks. On the RoboChallenge real-robot platform, it secures a specialist average success rate of 62 percent and a generalist average success rate of 37.3 percent. These results surpass prior state-of-the-art models including GigaBrain-0.1 and pi0.5, demonstrating the effectiveness of the embodied-native design. The complete framework is available at https://github.com/dexmal/dexbotic.

The DM0-base model from Dexmal supplies an initialization checkpoint for the training of the DM0 vision-language-action model. The model is a vision-language-action model with a base version. It accepts triple views of images as input and features an action dimension of 32. The model has a size of 2.4 billion parameters. It provides 3 billion parameters in total. The model uses a tensor type of F32 and BF16. The model employs a chat template. The model is distributed under the CC license.

Dexora模型地址：https://github.com/ZZongzheng0918/Dexora

OMPL模型地址：https://ompl.kavrakilab.org

硬件与算法工具——实现落地“最后一公里”

Beyond the brain, open-source initiatives for the "cerebellum" (motion control) and the "body" (hardware platforms) prove equally crucial.

在 ICRA 2026 上，乐聚机器人直接开源了物理身体的标准化基座。
其人形机器人平台的核心设计资源，包括完整 URDF 模型、25 自由度运动学参数及关节模组接口协议（支持 CANopen 与 EtherCAT 双栈），面向全球开发者全面开放，让算法团队无需从零进行机械设计即可部署 VLA 策略。

同步开源的还有配套高保真仿真环境及超2000条全身动态控制轨迹库，其中涵盖行走、转身以及双臂协同操作。这些资源专门弥合了人形机器人双足行走与全身协同操作的Sim2Real一致性难题。通过在硬件与仿真方面实现双双开源，乐聚机器人成功成为了本届大会硬件标准化生态的重要贡献者。

更关键的是其开源策略，即完整URDF模型公开发布，电机参数经工厂校准后直接写入URDF，从源头压缩Sim2Real鸿沟。该开源硬件的意义在于让没有千万级预算的高校实验室也能真刀真枪地研究人形机器人。

In the algorithmic realm, extreme sparse depth completion models stand as the ultimate challenge. These models utilize only ten point-level sparse depth measurements per frame to predict high-precision dense depth maps. On NVIDIA Jetson hardware, they achieve end-to-end latencies of 53 milliseconds. This capability enables low-compute, low-cost edge devices to possess high-quality 3D perception abilities, delivering a qualitative leap for drones and lightweight robots. Both the code and the outdoor environment test dataset have been open-sourced.

Figure 6: WideDepth establishes a benchmark for wide-field-of-view depth estimation through real indoor scenes and laser point clouds.

其他值得关注的算法包括SURE半密集不确定性细化特征匹配框架、MinNav基于光流的微型无人机极简导航方法以及LLM-TALE利用大语言模型引导强化学习探索的框架。

以下为硬件与算法工具完整清单：

“夸父”KUAVO人形机器人获取地址：https://github.com/LejuRobotics/kuavo-ros-opensource

乐聚全身动态仿真策略库的获取地址位于 https://github.com/LejuRobotics/kuavo-ros-opensource。

Mini Pi Plus获取地址：

In ICRA 2026, the Mini Pi plus humanoid robot platform from High Torque Robotics made its global debut. This platform stands at 75.6 centimeters tall and weighs only 15 kilograms. It therefore falls well below the 19.5 to 29 kilogram range of mainstream humanoid platforms. The full body features 23 or 27 degrees of freedom and a peak torque of 21 Nm. This configuration supports dynamic actions such as walking, running, and even backflips. The key open-source strategy lies in the complete URDF model that is released publicly. The motor parameters undergo factory calibration and are written directly into the URDF. This approach compresses the Sim2Real gap from the source. The significance of this open-source hardware lies in enabling university laboratories without multi-million-dollar budgets to conduct hands-on research on humanoid robots.

High-end powered by Mini Pi Plus debuted globally at ICRA 2026. The entire machine stands at a height of 75.6 centimeters and weighs only 15 kilograms, far below the 19.5-29 kilogram range of mainstream humanoid platforms. The full body features 23 or 27 degrees of freedom and a peak torque of 21 Nm, supporting dynamic actions such as walking, running and even backflips. More critically, the open-source strategy releases the complete URDF model publicly, with motor parameters written directly into the URDF after factory calibration. This approach compresses the Sim2Real gap from the source. The significance of this open-source hardware lies in enabling university laboratories without multi-million-dollar budgets to conduct hands-on research on humanoid robots.

精灵G2获取地址：https://www.agibot.com/products/G2

极端稀疏深度补全获取地址：https://github.com/ntnu-frl/depth-completion

MinNav获取地址：https://pear.wpi.edu/research/minnav.html

SURE获取地址：https://github.com/LSC-ALAN/SURE

LLM-TALE获取地址：https://github.com/llm-tale/llm_tale

效果验证——当526支战队共用一套开源工具链

前四章拆解了上半年开源生态的各个拼图。但一个核心问题有待回答：这些开源工具，到底能不能用、好不好用。

本次ICRA的 AGIBOT WORLD CHALLENGE 提供了一次绝佳的验证机会。赛事吸引了来自27个国家及地区的526支战队参赛，这些战队都统一基于 AGIBOT WORLD开源数据集、Genie Sim 3.0仿真平台以及精灵G2真机硬件完成了算法开发与评测。

The pioneering mechanism of online simulation evaluation combined with offline real-world finals establishes unified scoring criteria encompassing automated evaluation, real-machine stability, and Sim2Real transfer effects. This approach fundamentally represents a large-scale public stress test of the open-source toolchains. The fact that 526 teams competed on the same open-source system and produced reproducible results demonstrates the strongest proof of the ease of use and reliability of this toolchains.

开源社区的数据同样佐证了这一趋势。中兴RealMirror的120GB开源资产下载量突破2400+；原力灵机Dexbotic框架已服务清华、北大、普林斯顿等数十家院校以及腾讯等企业，汇聚超千位开发者；智元Genie Sim 3.0上线后迅速成为具身智能仿真领域的热门开源项目。

这些数字表明，开源工具并非仅作为论文附件的象征性发布，而是已成为真正被下载、使用并通过实际部署与验证的可用资源。

The barrier to developing embodied intelligence is being systematically lowered as open-source efforts transition from isolated breakthroughs to end-to-end coverage and as thousands of developers collaborate on the iteration of a single code repository. This is not a game played by a few giants but the foundation building for the entire industry.

来源：具身最强开源合辑汇总：从数据集到仿真平台，一文覆盖具身智能全链路 | 具身研习社

声明：本文来自具身研习社，版权归作者所有。文章内容仅代表作者独立观点，不代表爱力方立场，转载目的在于传递更多信息。如有侵权，请联系 copyright#agent.ren。