通用机器人操作系统 赋予具身载体无限的运动本能

强化学习、动作迁移、Sim2Real,领先的运动控制技术与跨本体能力

合作咨询

通用机器人OS
基础设施研发商

桥介数物是行业领先的通用机器人操作系统研发商。公司以运动控制为切入点,在强化学习、仿真训练、Sim2Real 及跨本体迁移等领域拥有深厚的技术沉淀与工程实践,致力于为行业提供安全、高效、可规模化的机器人运动能力。围绕核心业务,桥介数物已构建起由通用运动控制模型、跨本体全身运动数据工厂与通用机器人运动能力开发平台 RoboCraft AI组成的三位一体基础设施,打通“上下协同、全链赋能”的赋能路径,与产业链伙伴共筑开放、共赢、持续进化的机器人新生态。

40多家深度合作伙伴

具身智能明星创业公司、新能源头部车企、传统家电巨头的首选

70+

跨本体机型

截止2026年Q4,桥介数物运控模型已部署70多款不同的机型。

41%

2024 WRC

参展的27家人形机器人企业中,有11家使用桥介数物的运控方案。

39%

2024 WAIC

参展的18家人形机器人企业中,有7家使用桥介数物的运控方案。

最新消息

所有帖子

投资机构

桥介数物的高速成长与商业价值已获得国内多家顶级投资机构的认可

深创投 隐山资本 沂景资本 深圳天使母基金 奇绩创坛 复星创富 正轩投资 瑞丞基金 钧石创投 明荟致远 中移和创

快速构建运控能力

赋能第三方本体高效稳定运控能力,大幅缩短研发周期,全面实现降本增效

合作咨询

融合强化学习、模仿学习等先进运动学习技术
结合专家级的 Sim2Real 工程经验

通用机器人“小脑”
运动控制解决方案

行业领先的机器人通用运动控制模型

模型支持双足、四足、轮足等多构型机器人的统一训练与跨本体部署,有效解决不同机器人在动力学特性、硬件配置与控制策略上的适配难题,赋予机器人高物理一致性、强环境适应性与跨平台泛化性的运动能力,为客户提供安全、稳定、高效、可规模化部署的运控解决方案。

立即联系

通用全身控制器方案

标准化、高兼容性的通用运动控制解决方案,支持多种动作输入与跨本体适配,实现运动能力的快速接入、灵活拓展与高效部署。

  • 多源动作兼容,实现即接即用

    兼容大模型、动捕设备、仿真轨迹、动画关键帧等多种上层动作输入。通过统一接口,实现不同来源的动作指令快速接入与执行。

  • 动作灵活拓展,无需重复训练

    依托通用运动控制模型,支持执行未经针对性训练的新动作,具备较强的动作泛化能力,减少逐个动作训练和调试的工作量。

  • 跨本体快速适配,降低开发成本

    支持不同尺寸、关节构型的人形机器人,通过统一控制框架与适配机制,减少针对不同硬件重复训练的工作量,缩短研发与部署周期。

全向移动运控方案

聚焦机器人核心移动能力,融合强化学习与环境感知技术,实现自然、稳定、灵活的全向移动,提升复杂地形下的通行能力。

  • 强化学习驱动,运动自然稳定

    基于强化学习与AMP(对抗性运动先验)算法,结合物理约束与拟人/物运动特征进行训练,提升机器人行走的自然度、连贯性与稳定性。

  • 全向灵活移动,步态流畅自然

    支持前进、后退、左右平移、原地转向等多方向移动,实现自然流畅的拟人/物步态与平滑的姿态切换,满足不同场景下的灵活移动需求。

  • 环境感知融合,适应复杂地形

    结合深度相机等视觉感知技术,识别地形起伏与障碍物,动态调整运动策略,提升机器人在不平整、低摩擦等复杂地形下的稳定性与适应能力,并支持跨本体适配部署。

全身模仿运控方案

聚焦复杂动作的高精度模仿与追踪,让机器人精准复现复杂动作,满足高动态表演与个性化动作开发需求。

  • 深度模仿学习,动作精准稳定

    基于强化学习与DeepMimic(深度模仿)算法,学习复杂动作的运动规律与全身关节协调方式,实现高动态动作的精准追踪与稳定执行。

  • 复杂动作追踪,全身协调流畅

    支持舞蹈、功夫及高难度姿态等复杂动作的全身模仿与追踪,兼顾动作还原度、身体协调性与运动稳定性,提升机器人的动作表现力。

  • 跨机型灵活适配,支持个性定制

    针对不同尺寸、关节构型的人形机器人进行运动控制适配,结合动作重定向与模型优化,支持个性化动作开发与定制,满足不同机型及应用场景的需求。

感知与移动操作方案

融合运控、感知与操作能力,让机器人能够自主识别复杂环境、灵活调整运动方式,并在移动过程中完成操作任务,满足复杂场景下的作业需求。

  • 多模态融合感知,精准识别环境

    融合三维视觉点云与雷达数据,实时感知楼梯、障碍物及地形变化,为机器人自主规划通行路径和调整运动策略提供支持。

  • 多姿态灵活移动,适应复杂地形

    支持四足机器狗切换双足运动模式,以及“小恐龙”式直立行走等特殊步态,根据地形特点灵活调整运动姿态,提升各种机器人在不同高度、坡度及楼梯等复杂环境下的通行能力。

  • 移动操作协同,实现边走边作业

    融合足式移动底盘协同控制,在移动过程中兼顾身体平衡操作,提升动态抗扰与协调能力,支持机器人在复杂环境中完成移动操作任务。

快速构建运控能力

赋能第三方本体高效稳定运控能力,大幅缩短研发周期,全面实现降本增效

运控方案合作咨询

通用机器人
运动能力开发平台

机型适配、动作创作、仿真训练、真机部署、资产管理

立即体验

全面降低运动开发门槛

多年开发经验汇聚成标准化流程,大幅度降低运动控制开发门槛,实现高效交付

70+

跨本体项目积累

率先支持宇树G1、智元X2、加速进化T1等热门机型,打通自研硬件快速导入通路,真正打破软硬件壁垒。

90%

研发周期缩短

周期由 3 个月缩短至 1 周以内,提效超90%;开发团队由5人降至1人,极大缩减前期试错成本。

0

手工调试依赖

URDF校验与修复、动作数据预处理、强化学习训练、Sim2Real部署、参数调整等环节自动化处理。

0

本地安装部署

“WebAssembly + WebGL + MuJoCo + WebRTC”技术栈组合,无需安装任何本地插件或环境。

运动能力接入与应用开发

涵盖了新机器运动能力的快速构建以及成熟机型的动作能力开发和场景应用

AI校验

自动识别模型问题,全面输出可视化诊断报告

AI 校验诊断报告界面

URDF编辑器

可视化编辑关节模型,轻松解决诊断问题

URDF 编辑器界面

硬件检查

对齐电机、传感器、通讯参数,仿真即真机

硬件检查界面

运动能力训练

通用运动控制器,动作模仿、自由行走均可胜任

运动能力训练界面

多种机型

支持G1、X2、PM01、T1等多种主流机型,新机型可快速适配

G1 X2 PM01 T1

视频提取动作

支持从动作视频中提取高质量动作数据,同时也支持BVH上传

提取后的机器人动作

动作编排

丰富的动作资产可自由编排成个性化应用

动作编排界面

真机快速部署

3-5分钟完成真机连接到应用下发

真机部署界面

通用全身遥操作方案

支持 Xsens、诺亦腾、Pico等硬件设备,实现低延时高精度遥操作机器人,如影随形

跨本体

支持宇树G1、智元X2、众擎PM01等机型

低延时

及时响应,人机同频

高泛化

动作灵活拓展,无需重复训练

生态合作伙伴

与芯片、本体、场景等产业伙伴共建机器人生态

Intel
生态合作伙伴
京东
E-Bot Hub
芯擎科技 SiEngine
方特

体验 RoboCraft AI

快速为你的机器人构建一流且丰富的运动能力

立即注册

自研跨本体动作重定向算法

跨本体运动数据工厂

1000+小时/月·数据产能

跨本体全身运动数据工厂可将采集数据快速适配并分发至不同型号的机器人,实现高质量运动数据的高效复用,大幅提升数据生产与应用效率,加速运动控制模型的训练、验证与能力演进。数据工厂已形成每月1000+小时的数据产能,未来将分阶段向每月数万小时递进。

立即联系

自研跨本体数据引擎

自研重定向引擎打破传统“逐条动作调参、逐台机器人写专用解算”的限制。基于统一本体抽象层,结合高效流式数据管线与目标机器人的 URDF 拓扑定义,自动完成构型适配,将人体运动语义与拓扑关系统一映射至人形、四足、轮足等不同关节自由度与动力学构型。

多维数据增强体系

采用 RL 动力学后处理校正物理违背,消除浮空、穿透、滑移与力矩超限,实现从“运动学平滑”到“动力学可行”的升级。依托 3D 重建与仿真管线补齐深度、力学等视觉/力控模态,叠加环境扰动与 AI 语义标注,将单次采集指数级扩展为高密度训练样本。

双线自进化闭环

建立“设计-原始-重定向-增强”四级硬质检门槛,严格阻断污染数据进入训练集。同时,将模型训练侧的评估结果(失败画像)实时反向回写至上游采集与设计平台,形成“主动图谱覆盖 + 反馈精准补洞”的双线自进化闭环,确保数据资产随模型训练持续增殖。

CWM 数据工厂运行模式

采集、重定向、增强,再由训练结果反哺

主动覆盖

丰富运动多样性

  • 运动覆盖图谱
  • AI 原生管理平台
  • 多样性审阅与人员匹配

数据增强

物理 · 仿真 · AI 标注

  • 物理一致性后处理
  • 仿真多样性扩增
  • 语义标签与版本管理

同步采集

多源信息同步对齐

  • 人体动作 · 视频 · 场景同采
  • 统一时钟与数据契约
  • 自动质检,统一入库

跨本体重定向

适配多种构型的机器人

  • 统一本体抽象层
  • 自研 retarget 映射
  • 无需逐条调参,1000fps+

所有内容

媒体报道

桥介数物已获国内多家主流媒体的深度报道

返回

BridgeDP launches Cross-Embodiment Whole-Body Motion Data Factory, filling the data vacuum in motion control

BridgeDP has built a Cross-Embodiment Whole-Body Motion Data Factory, establishing a complete pipeline from motion design, synchronized capture, cross-embodiment retargeting, and data augmentation to training feedback

跨本体全身运动数据工厂

01 Introduction

We have built a Cross-Embodiment Whole-Body Motion Data Factory, establishing a complete pipeline from motion design, synchronized capture, cross-embodiment retargeting, and data augmentation to training feedback. As a continuously operating infrastructure, the data factory provides data fuel for humanoid robot operating systems, enabling the whole-body motion models within the system to continuously acquire cross-embodiment, trainable, and reusable data assets.

The improvement of model capabilities is increasingly constrained by data. With insufficient data scale, models struggle to cover enough motions; with unstable data quality, models learn incorrect contact relationships and body coordination; without cross-embodiment reusability, training assets must be re-produced whenever robot hardware changes.

For general-purpose whole-body motion models, data is no longer just training material but a critical asset that determines the boundaries of capability.

Based on this pilot and our engineering practice with various legged robots over the past two years, we are moving the data factory from internal validation to formal large-scale construction. It must address how to continuously plan motions, synchronously capture multi-source signals, perform cross-embodiment retargeting, conduct physical validation and data augmentation, and feed training results back into the next production cycle.

This article shares our phased thinking on the motion control data factory: what cross-embodiment whole-body motion data is, why we need a dedicated factory for it, and how the factory should operate internally.



02 What data do we need, starting from motion capabilities?

To answer "what data do we need," we must first answer "what motion capabilities do we want."

For general-purpose whole-body motion models, we want a motion capability that is upward compatible with multimodal action intentions, downward compatible with different embodiment hardware, safe and reliable, and able to continuously evolve in complex environments.

This capability places higher demands on data: models need data that simultaneously preserves whole-body coordination, task intent, contact relationships, environmental context, physical feasibility, and cross-embodiment reuse value.

However, existing data formats, when viewed individually, rarely satisfy these requirements naturally:


Motion capture data can accurately and structurally record human motion states but lacks environmental information and precise human-environment interactions;

teleoperation data is strictly tied to a specific robot embodiment, and its reuse value drops significantly when hardware changes;

first-person video focuses on end-effector and object interactions, failing to fully express the whole-body coordination among torso, lower limbs, center of mass, and contacts;

third-person video shows overall motion but makes it difficult to extract accurate and reasonable human motions. Each of these data types has value, but none alone can support the data loop required for general-purpose whole-body motion models.


Based on this judgment, we define data assets truly intended for training general-purpose whole-body motion models as Cross-Embodiment Whole-Body Motion Data (CWM), requiring CWM to satisfy at least the following four properties:


Cross-embodiment retargetability

: The same motion must be able to produce physically consistent training samples on multiple target embodiments with significant differences in link lengths, joint configurations, mass distribution, and actuation capabilities through a unified processing pipeline. This means the raw data itself must carry sufficient topology and kinematic information to support unified configuration mapping across different embodiments, rather than being locked into the joint space of a specific robot. Robot hardware will continue to iterate; if data serves only one generation of embodiments, it will depreciate along with that hardware. CWM ties data value to human whole-body motion semantics and transferable patterns, allowing a single dataset to be repeatedly utilized across multiple hardware generations.


Whole-body coverage

: Data must completely represent the torso, limbs, hands, fingers, and their coordination, rather than only preserving upper-body end-effector trajectories or lower-body gaits. Real tasks are often not simple concatenations of local actions; for example, "squat to pick up an object—hold it—turn and walk" involves lower-limb support, center-of-mass shift, torso posture, arm extension, finger grasping, and contact switching. Only by recording the coupling relationships among these body parts as a whole can the model learn the coordination patterns among locomotion, manipulation, and posture changes.


Physical feasibility

: A valid data sample must not only be kinematically smooth and reasonable but also be dynamically feasible on the target embodiment, without floating, penetration, slipping, instability, or torque limit violations. This is the hard threshold for upgrading CWM assets from candidate trajectories to training samples.


Multi-source augmentability

: CWM data simultaneously captures human motion, semantic labels, first-person video, third-person video, environmental assets, and object assets at the recording stage, endowing the motion with complete body, task, and scene context. Subsequently, we replay and augment the data in simulation, generating multi-view, multi-scene, and multi-physical-state training samples by customizing camera positions, replacing scene and object texture maps, and collecting whole-body contact forces and motion states.


CWM data satisfying these four properties cannot be obtained through simple capture, which is the starting point for building a cross-embodiment whole-body motion data factory.



03 Why build a data factory?

We have defined what CWM data is, but for model training, merely having "correct" data is not enough; data scale is equally crucial, a consensus already reached in the large model field.

Research on Generalist AI indicates that VLA models also exhibit clear data scaling laws; SONIC has systematically verified on humanoid robot whole-body motion tracking that expanding motion data volume significantly improves motion control capabilities. For whole-body motion control, this means data must cover not just a few standard motions but a large number of continuous motion combinations such as walking, turning, squatting, carrying, grasping, supporting, obstacle avoidance, balance recovery, and contact switching.

According to our internal assessment, training a truly general-purpose whole-body motion model ultimately requires hundreds of thousands of hours of high-quality CWM data; at this scale, small amounts of data have little long-term training value; what truly matters is the ability to continuously expand data scale.

At the same time, data diversity is equally important, because no amount of walking data can train a model that does backflips. The complexity of whole-body motion data lies not just in "more motions are better," but in having the correct data recipe and strict data quality control.

The model needs to see enough motion categories, contact states, task semantics, environmental variations, and target embodiment differences; at the same time, each data sample must undergo cleaning, annotation, retargeting, and physical validation. Otherwise, large-scale data can easily become large-scale noise. Problems such as foot slippage, body penetration, floating, instability, and torque violations are data pollution that directly degrades model quality, causing models to learn incorrect contact relationships, incorrect body coordination, and infeasible control patterns.

This standard also means external data cannot serve as the main source: public motion capture and online videos can supplement, but they are insufficient in quantity and quality to support training of general-purpose whole-body motion models.

Therefore, CWM data production must be designed as an industrialized production system, with capture as only one part. From the moment a motion is designed until it enters the training set, it must also undergo quality inspection, cross-embodiment retargeting, dynamics and simulation augmentation, semantic annotation, and a feedback loop from the model training side.

This production line must simultaneously define data recipes, production processes, and quality standards: which motions must be prioritized, which scenes and contact states are most scarce, which target embodiments need validation, which samples should be discarded, and which data yields the highest training returns—all of these need continuous tracking and feedback. The larger the data scale, the less we can rely on manual experience; the more general the model goals, the more we need reproducible, auditable, and iterable production processes.

This is the core value of the CWM data factory: using stable facilities, equipment, pipelines, professional teams, and quality inspection systems to turn general-purpose whole-body motion data into a sustainable production capability.

Professional motion designers define the motion taxonomy, the capture team handles high-quality synchronized recording, the engineering team manages cleaning, formatting, retargeting, and simulation replay, the algorithm team handles physical validation, training feedback, and data filtering, and the quality inspection team keeps unusable samples out of the training set.

Only such a factory-level system can continuously produce CWM data assets that are large enough, accurate enough, and clean enough, and that can be continuously updated alongside model training and robot iteration.



04 The data factory is not a "capture site" but an "infrastructure"

BridgeDP's Cross-Embodiment Whole-Body Motion Data Factory is a full-process infrastructure centered on producing CWM data assets.

It starts with motion design, clarifying motion categories, contact states, and task scenarios; during capture, it synchronously obtains multi-source data including human motion, video, contacts, environment, and objects; then through cross-embodiment retargeting, physical validation, and simulation augmentation, it transforms raw materials into trainable samples; finally, it uses training feedback to continuously revise the data recipe.


iwk0o1tylcVCUyU3DScmV37iwDk


Active coverage: enriching motion diversity

The first question the data factory must answer is "what to capture."A general-purpose whole-body motion model needs to see a continuously expanding motion space that covers the ways the body coordinates. This space cannot be just a pile of motion catalogs; it must be continuously filled along several mutually independent main axes:


  • Horizontal expansion of capability dimensions

: Capture plans must be organized by how the body is used, not by motion names. Basic dimensions such as locomotion, posture transitions, limb coordination, contact switching, and object manipulation form the foundation for subsequent complex capabilities. We care about how the body is invoked, how different body parts coordinate, and how center of mass and contacts change, rather than whether a specific motion has been captured.


  • Complex terrain, multi-person interaction, and environmental interaction

: These three types of scenarios are the most difficult beyond the basic dimensions and closest to real deployment needs, yet they are most easily overlooked and must be explicitly included in capture plans. Complex terrain changes support and foot placement strategies, multi-person interaction introduces rhythm alignment and spatial negotiation, and environmental interaction deeply couples body motion with objects, contact surfaces, and reachable spaces. They cannot be naturally extrapolated from flat-ground single-person motions and must be explicitly scheduled.


  • Subconscious behaviors and free expression

: Scripts can only define task boundaries; real motion contains many unwritten parts: individual movement habits, on-the-spot adjustments, and instinctive responses to unexpected events. Professional motion designers provide intent and constraints during recording while allowing performers space to complete motions in their own habitual ways, so that data covers both task goals and real body variations.


  • Motion recovery and failure fallback

: Whether a model can be deployed in real environments largely depends on whether it can stabilize during failures. Therefore, motion recovery must be separately included in capture plans, including rebalancing after imbalance, obstacle avoidance retraction after collisions, and recovery from falls or non-ideal postures. Such samples are typically scarce but directly relate to the model's safety boundary.


Capture diversity also needs explicit management at the source. The diversity of capture personnel and equipment directly affects the diversity and richness of CWM data: performers with different body shapes, ages, genders, and physiques bring differentiated motion postures, joint angle ranges, and center-of-mass control methods; different capture devices (inertial motion capture, optical motion capture, electromagnetic motion capture) differ in accuracy, coverage, wearing constraints, and applicable scenarios, which itself becomes a dimension of the data. Only by incorporating diversity of both personnel and equipment into capture plans can the model avoid learning only "the motion patterns of a certain type of person under a certain type of equipment."

These directions are organized and measured through a continuously updated motion coverage map, recording which combinations have been covered, which dimensions remain sparse, and which samples repeatedly fail after cross-embodiment transfer.

In addition to active coverage based on the map, the data factory explicitly receives data type requirement feedback from the model training side: which motion categories are unstable on which embodiments, which contact states yield the lowest training returns, and which samples passed quality inspection but brought no actual gains—all are translated into new data type requirements written back into the capture plan, so that "what to capture" is continuously calibrated by training results.

To truly convert the above requirements into executable capture tasks, we have built an AI-native data design and recording management platform within the factory, managing motion requirements, coverage map, scene assets, recording plans, data status, and training feedback in a single system.

The platform's core users are a group of full-time professional motion designers; they define motion semantics, break down body coordination, assess performability, and convert whole-body interaction, motion recovery, tool use, and scene tasks into recordable motion plans.

With built-in AI capabilities, the platform assists designers in three ways to produce motion plans:


In generating and expanding motion plans, the platform drafts motion descriptions and performs semantic-level generalization based on coverage map gaps and training feedback, deriving numerous variants along dimensions such as speed, body shape, and rhythm;

In visualizing plans, designers can choose to use AI to generate motion examples directly from text descriptions or keyframes, turning abstract descriptions into demonstrable reference motions;

In diversity review and personnel matching, the platform compares the current batch against the coverage map for distribution biases, alerts designers to over-collected or still-sparse dimensions, and assists in assigning each plan to the most suitable performer and capture equipment based on body shape, age, gender, and physique.


This toolchain closes the loop among the coverage map, designer judgment, and model training feedback within the same system, continuously converting "which motions are already learned stably, which motions have high transfer failure rates, and which scenarios still lack coverage" into production tasks that can be captured, tracked, and fed back.


Synchronized capture: multi-source information synchronized and aligned capture

Synchronized capture of CWM is not merely recording a segment of human motion; it must simultaneously answer four things within the same motion segment: motion intent, body movement patterns, interaction targets, and environment. "Whole-body" means that sub-tasks such as locomotion, manipulation, posture control, and contact changes all hold simultaneously within the same motion segment and cannot be reduced to simple concatenation of torso, hand, and leg trajectories. This inherently requires human motion, video, semantics, and scene to be recorded synchronously. Under current capture specifications, a complete record aims to synchronize the following four types of signals as much as possible, with specific availability depending on the capture scenario and target embodiment.


  • Human motion (BVH)

: The primary reference signal for cross-embodiment retargeting, carrying motion semantics, body coordination, center-of-mass changes, and posture transitions. We use different devices for different types of motions:

low-dynamic motions and motions on complex terrain are suitable for inertial motion capture, which is insensitive to site, occlusion, and terrain;

high-dynamic motions are suitable for optical motion capture or higher-precision optical-inertial hybrid devices, which can stabilize joint positions during rapid movements;

fine motions of the hands (grasping, tool operation, button pressing, twisting) are suitable for electromagnetic motion capture, which provides high-precision poses in small spaces.


  • Raw video

: Not directly entering the retargeting process, but a high-value auxiliary signal in the data factory: it supports video-based motion completion and human motion extraction, allowing massive internet videos to be incorporated into training assets and preparing visual modalities for navigation and manipulation; it is also used for training SLAM and estimating contact states between humans and objects. Equipment-wise, head-mounted cameras and external RGB/RGB-D cameras capture in parallel, providing first-person and third-person perspectives respectively.


  • Scene interaction assets

: Provide the environmental and object context in which motion occurs, a prerequisite for placing motion into simulation.

We collect two types: one is terrain and scene assets—room structures, ground undulations, fixed furniture—which determine reachable spaces and contact surfaces; the other is interactable object assets—objects that are carried, pushed, pulled, or used—which determine target geometry for manipulation tasks.

Technically, we use 3D Gaussian splatting + Mesh extraction for overall reconstruction, and for objects requiring precise poses, we further use optical Marker labeling. After entering simulation, these assets support reinforcement learning training and model evaluation.


  • Semantic labels

: Generated collaboratively by professional motion designers, on-site recorders, and AI annotation systems, defining motion boundaries, motion categories, scenes, and intent, determining how each sample enters the training set and how it is sampled, weighted, and evaluated during training.


The reason synchronization is necessary is that the value of whole-body motion lies not in any single modality but in the correspondences among different modalities. For the same "squat to pick up an object" motion, human BVH only shows how body posture changes; video shows where the object is and whether the hand actually makes contact; scene assets show the environment and interactive surfaces; semantic labels show motion boundaries and task intent. If these signals are not aligned, we cannot determine which frame's object contact corresponds to a hand trajectory, nor whether foot force matches the current posture, nor verify whether this motion can truly enter the training set.

To this end, the data factory establishes a unified capture clock and timestamp system for all capture devices: all devices complete spatial calibration and time synchronization before capture; during capture, the master control system uniformly manages task IDs, motion IDs, device status, and start/end signals; devices that support hardware synchronization prioritize using trigger signals, frame sync, timecode, or PTP for alignment, while devices that cannot hardware-sync record high-precision local timestamps and perform time synchronization correction through synchronized motions, calibration events, or post-processing algorithms.

After synchronization, each data sample must be organized into an asset that can directly enter the downstream pipeline; this work is also handled by the aforementioned recording management platform.

The platform performs on-site automatic quality inspection—checking time synchronization, calibration, trajectory completeness, bone length stability, keypoint anomalies, and motion segment boundaries, with AI assisting in checking motion semantics, performance consistency, and obvious recording anomalies—and unified ingestion—packaging all modalities under the same motion segment into a unified data package, binding session, device status, calibration version, time offset, frame loss, and quality inspection results, and completing alignment, resampling, and slicing based on the master clock, forming a minimal data contract that can directly enter retargeting and training pipelines.


Cross-embodiment retargeting: retargeting to robots of multiple configurations

The core solution to heterogeneity is motion retargeting: converting a motion in the coordinate system of a human or a reference embodiment into trajectories on a target robot embodiment. In industrialized production, the difficulty is no longer just "can we transfer a motion to a robot," but whether we can do so continuously, stably, and at low cost across a large number of motions and embodiments.

At the algorithm level, our self-developed retargeting engine targets "any motion × any robot model × any terrain."The input side covers arbitrary motions, upper-body/lower-body/whole-body, can process offline motion capture files and real-time motion capture streams, and supports motion signals from different sources such as video; the output side covers legged, humanoid, upper-limb, and composite-configuration robots with significant differences in structure, joint configuration, scale, and actuation capabilities, and can incorporate terrain constraints such as flat ground, slopes, stairs, and uneven ground into unified solving, without needing a separate set of dedicated solving logic for each motion, each robot, or each terrain type. The solver, based on kinematic solving and geometric constraints, integrates contact states, support relationships, spatial constraints, terrain constraints, joint limits, and body interaction relationships into a single solving process, outputting candidate trajectories that are semantically consistent, structurally reachable, and stably high-quality.

At the engineering level, it has three advantages that directly serve factory production.

First, no per-item parameter tuning and no motion templates: the cross-embodiment capability comes from a unified embodiment abstraction layer—when a new robot is integrated, we rely only on its URDF definition; the algorithm can automatically and quickly adapt to multiple configurations on this abstraction layer, without writing dedicated solving logic for each motion or robot, and without manual fine-tuning per motion.

Second, dual modes of streaming and offline: it can digest real-time motion streams from the capture side and batch-process existing motion libraries; this makes retargeting no longer an offline step of "process after capture," but enables retargeting while capturing—as soon as a motion is recorded, candidate trajectories on the target embodiment are already available, and quality inspection and subsequent dynamics augmentation can be connected immediately. In streaming mode, our retargeting tool supports output data from various devices such as Noitom and Xsens.

Third, cross-platform stable distribution: from engineering sites, capture scenes, and training clusters to target robot sides, it can be deployed and replayed consistently, ensuring the motion flow is based on the same algorithm implementation throughout the production chain.

At the capacity level, it is already the factory's main production service. Under current statistics, this retargeting algorithm can exceed 1000 frames per second on a single CPU core, about more than ten times the typical recording frame rate; we have prepared a computing cluster for this pipeline to continuously digest motion streams from the capture side and support parallel distribution of the same motion to robots of multiple configurations. In production terms, it compresses the hidden cost of "manual adaptation per motion" into a one-time engineering calibration when integrating a new embodiment, and reduces the "capture → retargeting → candidate training sample" pipeline time from days to near real-time.

The same human dance motion is retargeted across embodiments and mapped to multiple different robot bodies; some classified embodiments have been blurred. This effect is achieved without manual parameter adjustment or additional configuration of the algorithm.


Data augmentation: dynamics, simulation, and AI annotation augmentation

Cross-embodiment retargeting outputs high-quality candidate trajectories, but candidate trajectories are not yet final training assets. Data augmentation continues to transform these candidate trajectories into data that is more verifiable, more trainable, and easier for models to consume. We proceed along three main lines: dynamics augmentation, simulation diversity augmentation, and semantic annotation.

Dynamics augmentation places the most valuable, most difficult, and most physically demanding samples into the target embodiment's dynamics and contact models, using RL dynamics post-processing to simultaneously control tracking error and physical violations, upgrading candidate trajectories from "kinematically resembling" to "trackable on the target embodiment without penetration, torque violations, or friction cone violations." Samples judged infeasible enter quality feedback with specific failure reasons, rather than being discarded outright.

Simulation diversity augmentation repeatedly executes the same motion in different virtual environments, exponentially amplifying the coverage density of CWM assets.

On one hand, it fills missing modalities: through physical simulation and rendering pipelines, it supplements samples originally captured with only motion and video with modalities such as mechanical signals, depth maps, semantic segmentation, and multi-view images that were not originally captured;

on the other hand, it expands visual and scene diversity: replacing texture assets of objects and environments, adjusting materials and lighting, changing room layouts, introducing new interactive objects and initial states, and applying external force perturbations of different directions and intensities. The same motion can generate a large number of new samples across multiple target embodiments, multiple scene sets, multiple lighting conditions, and multiple perturbation conditions, so that the model sees not "one way to do this motion" but "a distribution of ways to do this motion."

Semantic annotation turns data into assets that can be retrieved, weighted, filtered, and reused by training pipelines. The AI annotation system assists in generating labels such as motion slices, motion categories, contact states, scene objects, task semantics, failure reasons, and capability dimensions, while professional motion designers review semantic boundaries and key samples, converging annotation outputs into a standard format usable for training sampling and evaluation bucketing.

The three types of augmentation share the same version and provenance record: each augmented sample is marked with which original motion it came from, which target embodiment, which dynamics post-processing, which simulation augmentation round, which annotation version, and whether it passed physical validation. This allows the training system to safely reuse, compare, and roll back augmented samples across different versions, and quality feedback can pinpoint specific augmentation stages when problems arise.

After cross-embodiment retargeting (green markers), dynamics augmentation produces dynamics-consistent high-quality data (red markers). Slippage, penetration, and floating are significantly reduced.


Quality feedback: feeding model training results back into the production system

Traditional motion capture quality inspection mostly checks whether trajectories are clean; the CWM data factory's quality management takes two steps: first, layered inspection along the production chain; then closed-loop feedback using model training results.

The first step is layered inspection. A sample, from motion requirements to the training set, must sequentially pass four independent quality inspections. Together, these four layers filter a candidate sample into an asset that can enter the training set, but whether it can truly train general-purpose whole-body motion capabilities can only be told by the model.

  • Design layer:

whether the motion requirement truly aligns with capability gaps, covers still-sparse cells in the motion coverage map, and can be translated into on-site executable motion plans. This layer controls "should it be captured."

  • Raw data layer:

whether the performer fully expressed the design intent, whether capture was synchronized, calibration was in place, and whether there were basic recording issues such as frame loss/drift/keypoint anomalies/bone length instability. This layer controls "was it captured correctly."

  • Retargeted data layer:

whether candidate trajectories are structurally reachable on the target embodiment, joints are within limits, contact relationships hold, and motion semantics remain intact after retargeting. This layer controls "does it still hold after transferring to the target embodiment."

  • Augmented data layer:

whether after dynamics post-processing it remains trackable, without penetration, torque violations, or friction cone violations; whether simulation augmentation and semantic annotation carry correct version and provenance records. This layer controls "is it truly effective in the training set."

The second step is result closed-loop. The training side aggregates every model evaluation result—such as which motion categories are stably learned on which embodiments, which transfers failed, which contact states yielded the lowest training returns, and which samples passed all four layers of inspection but brought no actual gains—into a writable failure profile: on which embodiment, which motion category, which contact state, and under which training version the problem occurred, attributing the problem to design, raw capture, retargeting, or augmentation.

The failure profile is directly written back to every upstream layer: the design layer adjusts the motion coverage map priorities and recording plans accordingly; the raw data layer adjusts capture specifications, synchronization strategies, and on-site quality inspection thresholds; the retargeting layer iterates algorithm capabilities; the augmentation layer adjusts dynamics post-processing intensity, simulation diversity configurations, and annotation standards.

Combining the two steps, the data factory forms a continuously iterating closed loop. In actual operation, it advances along two parallel lines: one line is active coverage based on long-term judgment, continuously expanding the human whole-body motion library according to the motion coverage map; the other line is feedback gap-filling from the model training side, backfilling every upstream layer according to the failure profile. With each round, the quality of data assets, cross-embodiment coverage density, and training returns all rise together: more accurate and faster with each round is the core source of compounding returns for the CWM data factory over time.


ohQ2xgXUT808I7xTGUgkaCWz64



05 Final words: the current status and future of our data factory

Over the past three months, we have run through the end-to-end chain of the cross-embodiment whole-body motion data factory in an internal pilot. The goal of this phase was not to pursue maximum capacity but to truly operationalize the entire production system: whether motion design can be systematically managed, whether multi-source capture can be stably aligned, whether retargeting can quickly adapt to new embodiments, whether augmentation and quality inspection can turn candidate trajectories into trainable assets, and whether training feedback can return to the next production cycle.

Along this chain, we have cumulatively produced nearly a thousand hours of high-quality CWM data; the whole-body motion model trained with this data ultimately completed key validation on over ten legged robots with significant differences in structure, actuation performance, mass distribution, and inertia distribution.

Now, this solution has completed internal feasibility validation, and the data factory is about to complete formal construction. The focus of the next phase is to shift from pilot validation to large-scale production—simultaneously expanding the site, capture studios, motion capture equipment, motion design team, performer roster, and algorithm/simulation/training computing clusters, so that the production line proven earlier can operate stably at a larger scale.

Our goal after the new factory is operational is to achieve a monthly output capacity of thousands of hours of high-quality CWM data for multi-configuration robots, and to climb in stages from "thousands of hours to tens of thousands of hours"; in this process, data quality, cross-embodiment reuse rate, and training gains will be continuously assessed under the same production standards, so that each new batch of data can answer how many embodiments it can run on and which motion categories it contributed real training gains to, rather than just "how many hours were captured."

“ 不仅是运动控制的底座构建者,更是跨越数字世界与物理现实鸿沟的探路人。”

关于桥介数物

桥介数物致力于打造具身智能时代的通用机器人操作系统
平台向上适配通用大模型与多模态感知指令,向下兼容异构硬件本体

加入我们

发展历程

桥介数物的发展历程及所获得的行业殊荣

人才汇聚

核心算法研究人员的硕士、博士背景占比达 80%
第三方足式机器人运动控制解决方案市场占有率第一

桥介数物打造了一支覆盖学术研究、工程研发与商业落地的全链条复合型团队,目前团队规模超过100人,其中研发人员占比超过80%。核心技术成员来自华中科技大学、南方科技大学、卡内基梅隆大学(CMU)等海内外知名高校,具备扎实的机器人控制、人工智能及系统工程背景;市场与商务团队核心成员曾就职于华为、中兴、京东等行业领先企业,拥有丰富的产业资源整合与商业化运营经验。

共登具身之巅

桥介数物诚邀全球精英与科研人才,用算法与物理法则,定义下一代机器人的运动轨迹

全球部署筹备中

加入 BridgeDP 桥介数物

定义下一代机器人的运动轨迹

15 个在招职位已开放,工作地点深圳。算法、工程、产品与商业岗位同步招聘
社招与实习都可以投递。

查看在招职位

在招职位

名单与飞书招聘站一致,目前开放 15 个职位,工作地点均为深圳。

合作对接

请填写您的核心需求与联系方式,我们的合作顾问
将在第一时间与您对接,为您提供专业的解决方案与技术评估。

1h 极速响应1v1 深度对接100% 商业保密

请填写您的姓名

请填写有效的电话或微信号

邮箱格式不正确

  • RoboCraft AI平台接入与应用开发
  • 运动控制解决方案合作
  • 跨本体运动数据工厂合作
  • 多维感知与规划
  • 投融资与媒体合作
  • 其他

官方电子邮箱

contact@bridgedp.com

24小时接收商业意愿函

商务顾问专线

+86 186 1001 0263

工作时间(9:00-18:00)

深圳总部

深圳市南山区留仙大道7013号航天创新大厦1801

欢迎预约考察合作

桥介数物微信二维码

扫码添加微信号实时解答