您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语： 产能公式从“人×小时”，变成“环境配置数×并行算力”。
作者丨幸丽娟
编辑丨岑 峰
训练一个可用的具身智能模型，需要千万小时的多模态交互数据；而现实仅有百万小时，剔除低质部分后，真正可用的更少。
缺数据，是具身智能行业的共识。
但只盯着“数据”这两个字，会错过更关键的问题：
一台机器人在固定货架前抓取，成功率可以很高。货架挪个方向，目标前多放一排饮料，换一侧伸手，成功率就往下掉。练过的，它能一遍遍做对；没练过的，它连手该从哪边伸都不知道。
行业缺的不只是数据量，还有“数据里的变化”。
过去两年，补缺口的主流办法是建场地、加采集人手。
据中国信息通信研究院，截至今年6月底，全国共建成启用超过70家训练场，另有40 余家在建或规划。
然而，大量训练场集中在叠衣服、收拾桌面、分拣搬运这类易展示任务上，复杂长程任务挖掘严重不足。数据小时数在涨，模型见过的空间布局没涨。
问题出在扩容的对象上。加场地，加采集人手，扩的是同一批场景的采集次数——货架还是那个货架，只是多采了几遍。
多样性要涨，靠的是条件变多：布局、遮挡、物体组合、接近路径，每多一种都是一次重新排列。
而一个训练场里能有多少种排列，取决于它被布置成什么样；改一次布置，采集流程就得跟着重走一遍。采集的人力、时间成本数倍往上叠。
产能越加越大，多样性原地踏步。行业真正卡住的地方，不是“采得不够多”，而是“变一次太贵”。
来自空间智能赛道的影溯，给出了另一种解法：不只采得多，还要穷尽空间的各种排列组合，采得“不一样”。
10月29日，影溯在杭州正式落地全国首家以仿真为主的具身智能数采中心，目标将产能拉升至月产百万小时级数据。
它以三维场景重建与生成为核心，把真实空间、生成式模型与仿真能力接到一起，做一套可编辑、可复用、可扩展的环境与数据供给体系。
在“真实数据为王”的共识里打出“仿真”的旗号，需要底气。
这种底气，源自影溯直接变革数采方式，同时解开数据多样性和产能两大“卡点”；更源自，它自己作为一家空间智能基模公司，最知道模型需要什么数据，以及什么才叫真正的高质量数据。
01
数据产能和多样性，
卡在“每加一种变化就重来一遍”
要看懂影溯的选择，得先看清楚它要替代的数采流程，成本到底花在哪。
EGO第一视角采集与遥操作，是具身智能获取真实行为数据的两种主要方式。EGO需要人员进入环境、完成操作；遥操需要操作者持续参与，并适应具体设备和任务。两者方式不同，但都绕不开“人”。
绕不开人，就绕不开组织。新增一种空间布局，补充一类操作条件，流程就得再走一遍：找场地，约档期，把人凑齐，设备调好，采完拆掉。
此外，采集人员的任务熟练度、操作习惯，以及对不同设备的适配程度，也会直接影响采集效率和数据一致性。
重组流程不仅繁琐，而且还没法并行。场地要一个一个谈，人要凑齐了才能开工，采完的场景拆掉就没了。
于是每增加一种变化，就要重付一次全额的组织成本。
这就是多样性涨不上去的真正原因：不是行业不想加条件，是每加一种，就得从头组织一遍，成本线性叠加。
影溯要做的，就是把这个成本降下去，同时把数据多样性拉上来。
在仿真环境里，空间布局、物体位置、遮挡关系和任务条件可以通过生成、编辑与组合来扩展，不必为每一种变化重新安排实体场地和人员操作。同一项任务可以被放进不同的环境条件里，覆盖那些真实采集中很难穷尽的变化。
对算法团队来说，它是扩展训练分布、复现失败条件的空间；对机器人本体厂商而言，它是一套能够反复配置和验证的环境基础。
数据多样性与产能的同步提升，还来自数据生产全流程的并行计算。
基于少量真机或遥操示范种子，系统结合运动规划、策略驱动与自动化物理质检，在数百个并行环境中自动批量合成对齐了 RGB、深度、分割、6D 位姿与动作指令的闭环轨迹。环境能重复使用，条件能批量调整，数据生产的规模由此随算力和任务系统扩展，而不再只随人数线性增长。
两种范式的分野就在这里：过去两年行业比的是谁的场地多、采集员多，产能公式是“人×小时”，小时数在涨，条件空间并没有涨；影溯数采中心的公式是“环境配置数×并行算力”，每多一种配置，就多一组条件覆盖，多样性跟着算力一起扩。
基于这套体系，影溯数采中心有望实现月产百万小时（或日产十万小时）的数据生产能力——这意味着单中心短期运转的有效产出，就能跨越全行业过去两年攒下的数百万小时存量总和。
而比这些数字更关键的，是它指向的工作方式的革新：模型在哪类遮挡、布局或操作条件下出问题，就针对这些条件扩展环境、补充样本、开展回归测试，把从发现问题到再次验证的周期压短。
数据生产与研发贴得越紧，“如何生产出高质量高需求数据”才有了方向可依，也能少走“采了大量、却用不上”的弯路。
02
真实空间定准，仿真生成扩量
但在具身智能行业，仿真一直有个绕不过去的质疑：环境不够真。
这是Sim-to-Real 的老问题，也是“真实数据为王”这句话的底气所在。
影溯没有回避，而是把问题拆开：几何、外观、空间关系交给真实采集去锚定，扩展和重复验证交给仿真去做。
于是场景资产有了两条进料管线。
一条从真实空间来。
通过与禾赛Kosmo的深度合作，影溯把采集侧的硬件与重建能力，同自己模型侧的3D升维及物理仿真能力衔接起来。
Kosmo这套三维世界数据引擎的厉害之处，在于它把“真实场景变成可计算三维资产”的时间成本和精度门槛同时压了下来：普通室内100平方米约15分钟完成扫描，3DGS重建达到毫米级，一次输出3DGS、mesh、depth三类对齐数据。
但硬件扫描直接输出的是一层粘连的“静态视觉壳体”，物体无法单独抓取、遮挡背面存在空洞。Kosmo 负责把物理空间数字化，而影溯的核心壁垒，是通过 3D 升维与世界模型把“静态壳体”拆解并赋予物理属性，变成机器人真正能抓、能碰的“可交互资产”。
也就是说，Kosmo采集的真实世界数据，为场景几何、外观和空间关系提供了可靠参照；而影溯则把这些场景数据，变成可编辑、可复用、可扩展的仿真环境资产。真实采集、三维重建、世界模型训练，这条链路得以完整打通。
另一条靠生成式模型扩展供给。
影溯推出的世界模拟器 Topos Pro，核心能力是把任意空间输入变成可理解、可编辑、可交互、可仿真的对象级世界。
它支持单图、多图、视频、LiDAR与无人机航拍等多源输入，自动完成场景物体的分割、与空间几何的对齐、以及物理属性赋予。更关键的是，其中的桌子、椅子、墙面、设备可以被单独选择、替换、调整，并赋予碰撞、可移动、可接触等属性。
视频再多，模型也只能看；能被行动、接触和验证的世界，才是具身训练真正缺的东西。而 Topos Pro做的，就是把“看”变成“可交互”、“可编辑”。
两条管线分工清楚：真实场景重建负责定住“像不像”，生成能力负责补上“够不够多”。一次采集的素材不再只用一次，互联网上的图片、视频也能被升维成可用的三维场景。
不同于其他数采方式“采一次、用一次、改一次布置重走一遍流程”；影溯利用真实采集定锚点， Topos Pro 把同一次采集重组成无数种布局、遮挡和物体组合，每一次重组都是一次新的条件覆盖。
场景资产由此从“消耗品”变成“可复用的原料”：采一次，不只用于一次，而是长出一片条件空间，可以无限循环利用。
03
一次场景输入，
多轮环境编辑与验证
场景数据到位之后，还有一道关键处理：让场景里的东西能被单独操作。这一步，是场景能不能从“采一次就拆掉的消耗品”变成“可反复配置的资产”的分水岭。
以“从货架取出指定商品，再放入周转箱”为例。系统需要保留货架、商品和容器之间的空间关系，把目标商品独立出来，补全被遮挡的部分，并修复提取后留下的背景空缺。
具体而言，系统会进行三步处理：
第一步，实例级三维解耦。 把目标商品从货架网格里剥离出来，成为能够独立选择、移动和配置的对象，同时保留它与货架、相邻物品之间的空间约束。精准解耦，是后续进行物体级再编辑的前提。
第二步，遮挡完形补全与三维背景修复。补全被遮挡的背面几何，实时修复商品移走后留下的背景空洞，让物体被移动或替换时，不会出现几何缺失。这一步，决定了场景是否能“经得起”反复编辑。
第三步，任务级物理属性标定。 按任务要求标定质量、质心与摩擦系数，精细处理物体底部、接触面及附近障碍物的几何，让仿真里的接触判断站得住。
经过这三步处理，研究者就可以改变商品位置、增加遮挡、调整放置区域，观察机器人在不同条件下的表现。一次场景输入，由此能够服务多轮环境编辑与任务实验。
影溯的技术优势，恰好体现在这些具体的能力之上：把多源输入转化为三维场景，让场景中的关键物体可以被编辑与交互，再结合目标机器人和任务进行配置与验证，使空间内容真正进入具身智能的研发流程。
这一步回应了行业对“仿真数据” Sim-to-Real 的质疑。仿真够不够真，是靠实实在在的结构解决的：几何、外观、空间关系来自真实采集，物理属性经过配置与标定，仿真在此之上做扩展和重复验证；但仿真不是终点，策略最终还要用真机测试检验。
换句话说就是，仿真并没有直接“替代”真实采集，而是真实采集建立的结构化场景理解被仿真“放大”了：放大到能覆盖长尾场景、能反复配置验证、能批量进入具身智能的研发流程。
04
按用途给环境“分层”，
让数据供给跟上研发迭代
资产能用起来，还得能被挑中。对算法厂商和机器人本体厂商来说，数采中心的价值最终要落到可使用的环境和数据上。
不同任务对环境的要求并不一样。影溯针对不同任务类型，定义了工业化的“三级环境交付标准”：
L1 移动导航级，交付大尺度空间拓扑、可通行区域与动态干扰，对应的是“能不能走过去”。
L2 灵巧抓取级，交付毫米级物体几何、碰撞体与密集遮挡组合，对应的是“手能不能伸进去、抓不抓得起来”。
L3 接触与铰接动力学级，针对开抽屉、插拔这类复杂交互，交付关节阻尼、接触力与动力学标定，对应的是“用了多大力、会不会被卡住”。
三级之间是精度与物理保真度的递增。这个区分有实际意义：行业里“高质量数据”长期是个形容词，各家都说自己产的是高质量，但没法验收。
三级分层之后，数据的几何精度达到哪一档、碰撞体如何定义、阻尼标定在什么范围，都能在交付单上明明白白地体现，下游团队也更容易挑到适合自身任务的资源。
还有一个更值得留意的地方：越往上一级，真实采集场景越难覆盖，仿真的相对优势反而越大。L1级靠实地扫描还能铺开量；到 L2级，密集遮挡的排列组合在实体场地里只能靠一次次重摆去凑；到 L3 级，每个抽屉的阻尼、每处接触的力学响应都不一样，几乎无法做到穷举。
因此，任务分层划出的不只是标准，也是仿真真正能发挥价值的那段区间。
场景与素材的复用，可以减少重复搭建环境的工作；持续补充布局、物体和任务条件，则能够支持模型测试、失败复现与策略比较。同时，客户的使用反馈，又可以成为下一轮数据生产的依据。
到这里，影溯实现了整个数据采集流程的闭环：真实采集提供参照，场景重建把它变成资产，生成与仿真扩展覆盖范围，任务验证给出反馈，研发反馈再指导下一轮生产。真实世界、生成模型与仿真能力在这个闭环里各司其职。
这一闭环的意义不仅仅在流程本身，还能进一步加速生产速度：数据生产一旦能接住研发反馈，迭代就不再是一轮一轮地等采集排期，而直接沉淀为“发现问题、扩展条件、重新验证”的连续动作。
05
开放生态共建：场景资源从哪来
说到这里，仿真数采中心的建设，还面临一个更现实的问题：环境再多，也得有足够的场景资源。
因此，场景资源、采集组织、技术能力与应用需求之间的协同，非常关键。这不是一家公司关起门能补齐的链条。
基于此，影溯选择与杭州每回互动数采科技有限公司合作。作为 A 股数据智能龙头每日互动生态孵化的具身数采专精力量，每回互动承接了每日互动 —— 国内含数率领先的数据智能上市公司，在规模化设备覆盖、DiOS 数据治理以及覆盖全国超 2900 个区县公共服务网络上的深厚积累。
也就是说，影溯拿到的不是一个现成的数据产能，而是一个能把下游需求前置到采集设计阶段的入口。
这个入口的价值不在交付端，而在设计端。每回互动的真实场景采集组织能力，同影溯的三维重建、生成及仿真能力连接起来，推动真实数据向可复用环境与任务素材转化。
下游需要验证什么任务、保留哪些空间关系、补充哪些环境变化，都可以更早进入采集与生成流程，而不是等到数据交付之后才发现问题。场景的“可编辑性”不再是“事后补救”，而是直接变成生产流程第一天的“输入条件”。
但一个伙伴接进来的场景，撑不起模型需要的广度。影溯更早就在想这件事。
早在今年8月，它就联合禾赛、宇树、云深处、商汤、华策影视等20多家产学研单位，共同启动了“大规模空间智能3D数据开放计划（SIDO）”，围绕空间数据与评测基准开展共建。
SIDO 要解决的不是“多采一些数据”，而是把数据供给的结构拆开重排。没有任何单一来源能同时满足广度、真实性和长尾覆盖，所以 SIDO 要探索的是三类互补的数据路径：
第一，真实世界采集。产业合作方通过激光雷达、多视角影像、机器人和其他传感器实地采集真实场景，管锚点。
第二，互联网内容升维。通过重建生成技术，把互联网公开的图片、视频升维为3D/4D场景，管广度
第三，仿真与合成。通过仿真合成现实中难以采集的极端与长尾场景，管泛化边界。
三条路径不是并列备选，而是互相咬合：缺了实采锚定，升维出来的数据没有物理参照；缺了升维扩量，实采撑不起多样性；缺了仿真补长尾，前两条加起来仍会漏掉最需要泛化的那部分条件。
这个结构，没有哪一家公司能独自补齐。影溯很早就认识到这一点，所以它牵头行业共建和合作，与此同时，更先一步把自己最擅长的仿真这条线做扎实，用可复用的仿真能力去补实采和升维都够不到的长尾条件，再把这条能力交进共建框架。
不仅如此。为了让规模始终对准模型的真实缺口，配套的评测 Benchmark体系反向定义标准：不是采得越多越好，而是用评测结果告诉下一轮该采什么。模型在哪类条件下失败，下一轮就补哪类数据；数据越多，越对准失败点，而不是越采越散。雷峰网(公众号：雷峰网)雷峰网雷峰网
当具身数采的产能公式从“人 × 小时”切换为“环境配置数 × 并行算力”，由真实采集定锚点、Topos Pro 世界模型扩场景、高保真仿真跑出规模化产出交互经验的闭环已经成形。
影溯在杭州落地的不仅是一座仿真数采中心，更是推动具身智能告别“手工作坊”、迈向通用泛化的算力引擎。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: The energy formula changes from " man-hours " to " environment configurations x parallel algorithms ".
The author is lucky to be here.
Edit Peak
Training an available smart model requires tens of millions of hours of multimodular interactive data, whereas reality is only a million hours and less is really available after removing the low-quality part.
The lack of data is a common understanding of the smart industry.
But looking at the word "data" alone would miss more critical questions:
success rate can be very high. The shelf moves in a direction, with an additional line of drinks in front of the target, and the other side stretches, and success rate falls down. It's practiced, it can do it over and over; it doesn't know where it should be.
The industry lacks more than just data volume, but also “changes in data”.
Over the past two years, the mainstream approach to filling the gaps has been to build sites and gather additional personnel.
According to the China Institute of Information and Communication Research, by the end of June of this year more than 70 training sites had been built and opened, and more than 40 others were under construction or planning.
However, a large number of training sites are concentrated on easy-to-exhibit tasks such as stacking of clothes, packing of desktops and sorting and removal, and complex long-range tasks are seriously under-excavated. The number of data hours is rising and the space layout that the model has seen has not increased.
The problem is with the outreach. The addition of a space and a pick-up capacity was extended to the number of pick-ups of the same scene — the shelf or the shelf — but only a few times.
Diversity increases by changing conditions: layout, shielding, clustering of objects, approach paths, and reordering of each.
And the number of rankings in a training field depends on what it is set up to do; a change in the layout requires that the collection process be focused. The collected manpower and time costs are multiplied.
The greater the capacity, the greater the diversity. The real scalding place for the industry is not “not enough to pick”, but “to make it too expensive”.
The hindsight from the space-intellectual track offers another solution: not only to take much more, but also to exhaust the various arrays of space and to extract “different”.
On 29 October, in Hangzhou, where the country ' s first self-improvement-based smart number mining centre was officially landed, the goal was to raise production capacity to 1 million hours per month.
It has a three-dimensional re-establishment and life at its core, bringing together real space, generation models and imitation capabilities to make a redactable, reusable, scalable environmental and data supply system.
In the consensus that "real data is king" the "fiction" flag needs a bottom line.
This background, which stems from the direct retrogression of the data captures, as well as from the release of the two “card points” of data diversity and capacity; and, more importantly, from its own role as a space-smart base model company, which knows best what data are needed for models and what is really high-quality data.
01
Data capacity and diversity,
It's stuck in "Every change goes over again."
Depending on the choice of phototrace, it is necessary to see what it is to replace with.
EGO's first perspective collection and remote operation are the two main ways in which intelligence can obtain real behavioural data. EGO requires personnel to enter the environment and complete its operations; telecommuting requires the continued involvement of operators and adaptation to specific equipment and tasks. Both approaches differ, but they do not bypass the term “human”.
If you can't get around, you can't get around. A new space layout, supplemented by one type of operating condition, would have to go through the process again: to find a space, a slot, to get people together, to set up the equipment and to remove it.
In addition, the job proficiency of collectors, their operating habits and their suitability for different equipment can have a direct impact on collection efficiency and data consistency.
Reorganization processes are not only cumbersome, they cannot go hand in hand. It takes one to talk, people to get together to start work, and the scenes that have been taken out will be destroyed.
For each additional change, the full cost of the organization would have to be recosted.
This is the real reason why diversity does not rise: it is not the industry that does not want to add conditionalities; it is every addition that it is organized from the top to the top and the cost is linear.
What we need to do is to lower this cost while bringing data diversity up.
In a simulation environment, space layout, object position, shielding relationships and mission conditions can be expanded by generating, editing and grouping, without the need to rearrange physical sites and people for each change. The same task can be placed in different environmental conditions, covering changes that are difficult to exhaust in real collection.
For the algorithm team, it is a space to expand the distribution of training and to repeat failure conditions; for robotics, it is an environmental foundation that can be configured and validated over time.
The simultaneous increase in data diversity and production capacity is also derived from parallel calculations of the full data production process.
Based on a small number of live machine or tele-drive demonstration seeds, the system, in combination with motor planning, strategy-driven and automated physical quality checks, has aligned the close loops of RGB, depth, partition, 6D position and action commands in hundreds of parallel environments. The environment can be reused and the conditions can be adjusted in bulk, so that the scale of data production can be expanded with calculus and mission systems, rather than with linear population growth.
This is the difference between the two paradigms: over the past two years, the industry has had more sites than its collectors, the energy formula is “person x hours”, the number of hours is rising, and the condition space has not increased; the formula for the backtracking centre is “environmental configurations x parallel calculus”, each of which is covered by more than one set of conditions, and diversity expands with calculation.
Based on this system, the IRC is expected to achieve a data production capacity of one million hours per month (or 100,000 hours per day) — which means that a single centre can function in a short-term, effective output that cuts across the millions of hours of stock that have been saved over the past two years across the industry.
What is more critical than these figures is the innovation of the way it works: in what kind of mask, layout or operating conditions the model is problematic, extending the environment to these conditions, supplementing sample, conducting regression tests, and reducing the cycle from finding problems to revalidation.
The closer data production and research and development is, the more the “how to produce high-quality, high-demand data” can be directed, and the less the “mass, but no use” curve can be taken.
02
Real space is set.
But in the smart industry, there's always a question about how to get around it: the environment isn't real enough.
This is the old problem of Sim-to-Real and the bottom line of the phrase "real data is king".
The problem is not avoided: geometry, appearance, spatial relations are handed over to real collection for anchoring, and extension and re-certification are given to simulations.
So the scene assets have two feed pipes.
One from real space.
Through a deep collaboration with Kosmo, the hardware on the side of the collection was traced back to the capacity for reconstruction and to the 3D-liter and physical simulation capacity on the side of the model itself.
The power of the Kosmo 3D World Data Engine is that it translates the "real scene into a time cost and precision threshold for calculating 3D assets" At the same time, there was pressure: approximately 15 minutes to scan 100 square metres in the normal room, 3 DGS reconstruction to millimetre level, and a single output of 3 DGS, mesh and decth alignment data.
However, the direct output of the hardware scan is an adhesive “static visual shell” where the object cannot individually capture and shield an empty back. Kosmo is responsible for digitizing the physical space, while the central barrier of hindsight is to disassemble the “static shell” and give physical properties to the 3D elevation and world model and turn it into a truly capable and accessible “interactive asset”.
In other words, the real world data collected by Kosmo provide a reliable reference to the geometry, appearance and spatial relationship of the scenes, while the hindsight transforms these scene data into editable, reusable, scalable environmental assets. Real collection, 3D reconstruction, world modelling training, the link is fully accessible.
The other one expands supply by means of a production model.
Topos Pro, a world simulator, with a core capability to make any space input an understandable, editable, interactive, replicable object-level world.
It supports multi-source inputs such as single graphs, multigraphs, videos, LiDAR and drones, auto-completes the partitioning of scene objects, alignment with spatial geometry, and physical attributes. More crucially, the tables, chairs, walls, equipment can be selected, replaced, adjusted and given characteristics such as collision, mobility, accessibility, etc.
As many videos as possible, models can only be seen; a world that is actionable, accessible and validated is what is really lacking in training. And what Topos Pro does is to turn "see" into "interactive," "editable."
There is a clear division between the two lines: the real re-establishment of the scene is responsible for “appearances”, and the generation capacity is responsible for “sufficient enough”. asset, which is collected at a time, is no longer used only once, and images and videos on the Internet can be upgraded to a three-dimensional scenario.
Unlike other digital methods, “take, use, reset and reroute the process once”; and use the real collection of anchors, Topos Pro forms the same collection into countless layouts, shields and combinations of objects, each of which is covered by new conditions.
As a result, the scene asset was changed from a “consumable” to a “reusable raw material”: it was taken once, not only once, but rather in a conditional space that could be recycled indefinitely.
03
One sceneial input,
Multicycle environment editing and authentication
Once the scene data is in place, there's a key process: to make the scenes work alone. This step is whether the scene can move from a “take-out-of-consumable” to a “repeated” watershed.
Take the example of “the removal of a specified commodity from the shelf and its placement in a swing box”. The system needs to preserve the spatial relationship between shelves, commodities and containers, to separate the target commodity, to complete the hidden parts and to repair the background gaps left by extraction.
Specifically, the system will be addressed in three steps:
First step, a three-dimensional solution at the instance level. The target commodity is removed from the shelf network and is the object of independent choice, movement and configuration, while retaining the space constraints between it and the shelf and the adjacent goods. Precision is a prerequisite for subsequent object-level re-editing.
Step two, cover the finish and three-dimensional background. Complete the hidden back geometry, repair in real time the background holes left behind by the removal of the goods, so that the objects are moved or replaced without geometric defects. This step determines whether the scene is “affordable” for repeated editing.
Step three, task-level physical properties. The mass, mass and friction coefficients are determined as required by the mission, and the geometry of the bottom of the object, the surface of exposure and the nearby barrier is finely processed so that the mimicry of exposure can stand.
Through these three steps, researchers can change the position of the commodity, increase the cover, adjust the area and observe the performance of robots under different conditions. One scenario input, which will serve multiple rounds of environmental editing and mission experiments.
Retroactive technological advantages are reflected in these specific capabilities: Translating multi-source input into a three-dimensional scenario allows key objects in the scene to be edited and interacted, then configured and validated in combination with target robots and missions, so that space content is actually entering a smart R&D process.
This step responds to industry's challenge to Sim-to-Real. The simulation is true enough to be solved by a solid structure: geometry, appearance, spatial relationships come from real collection, physical properties are configured and marked, and simulations are extended and repeated; but simulations are not the end point, and strategies are ultimately tested using real machines.
In other words, the simulation is not a direct “replacement” of real collection, but rather the structured scenario of real acquisition is understood as a “magnification” of real acquisition: amplification to cover long tailings, reconfiguring validation, and mass access to smart research and development processes.
04
"Stratification" of the environment by use,
Let the data supply keep up with R & D.
Assets can be used and selected. For algorithms and robotics manufacturers, the value of digital centres ultimately falls to the available environment and data.
The environmental requirements of different missions are not the same. The three-tier environmental delivery criteria for industrialization have been defined by reference to different types of tasks:
L1 Mobile navigational stages, delivering large-scale space expansion, accessible areas and dynamic interference, corresponding to “capability to pass”.
L2 Smart grab-level delivery of millimetre-grade objects geometry, collision body and dense shield combinations, corresponding to “capable hand in or uncapable of catching”.
L3 Exposure and reconnection dynamics level, for complex interactions such as opening drawers and liftings, delivery of joint blockages, contact forces and motors, corresponding to “how powerful it is and whether it is stuck”.
Between three levels is the increase in accuracy and physical authenticity. This distinction has practical meaning: “quality data” in the industry has long been an adjective, with all parties saying they produce high quality, but not acceptance check.
After the three tiers, the geometric accuracy of the data, the definition of the collision body, the extent of the deniability marker, can be clearly and explicitly reflected on the delivery order, and the downstream team can more easily identify the resources appropriate to its mandate.
There is also something more noteworthy: the higher the level, the more difficult it is to cover real collections, the greater the real comparative advantage. The L1-based field scans can also spread the amount; at the L2 level, the dense arrays can only be rearranged over and over again in the physical site; and at the L3 level, each drawer has different blockages, and the mechanics of each contact can hardly be exhaustive.
Thus, task stratification is not just a standard, but a simulation of what really works.
The re-use of the scene and asset can reduce the duplication of the construction of the environment, and the continuous replenishment of layouts, objects and mission conditions can support model testing, failure recurrence and strategic comparison. At the same time, client feedback can be used as a basis for the next round of data production.
Here, the entire data collection process is closed: real collection provides reference, scene reconstruction transforms it into an asset, generation and simulation of extended coverage, mission validation gives feedback and research and development feedback guides production. The real world, the capacity to generate models and imitations, works in this closed circle.
The significance of this closed circle is not only in the process itself, but can further accelerate production: Once data production is able to capture R & D feedback, it is no longer a round-of-the-square collection schedule, but it is a continuous process of “discovering problems, extending conditions, revalidating”.
05
Open Ecology Together: Where are the scene resources coming from?
At this point, the construction of analogue centres also faces a more realistic problem: the environment has to have enough resources for the scene.
Therefore, synergy between scene resources, collection organizations, technical capabilities and application needs is critical. This is not a chain that a company can fix by closing its doors.
On this basis, the hindsight chose to work with Hangzhou ' s Science and Technology Company Ltd. for every interaction. As a powerful force for daily interactive ecocubation of the A-unit data intelligence, daily interactions are carried over every time - the country ' s leading data intelligence listing company, which accumulates in size-based equipment coverage, diOS data governance and coverage of the country ' s over 2900 district public service networks.
That is, it is not an off-the-shelf data production, but an entry point that can pre-position downstream demand to the collection design phase.
The value of this entry is not at the delivery end, but at the design end. The organizational capacity of the real scenes of each interaction, linked to the three-dimensional reconstruction, generation and simulation of the hindsight, promotes the conversion of real data to a reusable environment and mission asset.
Downstream, which tasks to verify, which spatial relationships to retain, and which environmental changes to supplement, can enter the collection and generation process earlier, rather than wait until data are delivered. The "editability" of the scene is no longer a "remediation after the event" but is a "entry condition" on the first day of the production process.
But the scene of a partner coming in cannot sustain the breadth required by the model. I thought about it earlier.
As early as August this year, it launched the Large Space Intelligence 3D Data Open Program (SIDO) on the basis of spatial data and evaluation benchmarks, with more than 20 production research units, such as the United Wheats, Woo Trees, Deep Clouds, Business Soup and China Plan Film.
SIDO is not going to solve the problem of “excessing some data”, but rather to dismantle the structure of the data supply. There is no single source that can satisfy both breadth, authenticity and long end cover, so SIDO is exploring three complementary data paths:
First, the real world collects. Industry partners collect real scenes through laser radar, multi-perspective images, robots and other sensors in the field, and pipe anchors.
Second, the Internet content has increased. 3D/4D scenes, wide-scaled by re-establishing technology for Internet-open images and videos
Third, simulation and synthesis. Translating the border through extreme and long tailings that are difficult to capture in the simulation and synthesis reality.
The three paths are not a parallel option, but rather a combination of each other: the lack of solid anchors and the absence of a physical reference to the elevation data; the lack of a high-dimensional build-up, which does not support diversity; and the absence of a simulation, which, together, still leaves out the part of the first two that most needs to be generalized.
This structure, which none of the companies can fill alone. This was recognized long ago, so it led the industry to build and work together, and at the same time, first, to make the line of its best imitation solid, using reusable imitation capabilities to fill long-term conditions where neither the extraction nor the advancement of the force would be sufficient, and then to put that capacity into a common framework.
More than that. In order to keep scale up to the real gaps in the model, the accompanying assessment of the Benchmark system is a reverse definition: not as much as possible, but rather what to tell the next round. Under what conditions the model fails, and the next round adds the type of data; the more the data are, the more the failure point, the more the failure point is, and not the more it becomes. Thunder Peak Mine Peak Network
When the power formula with body size is changed from "human x hour" to "environmental configuration x parallel computing power", the closed loop of real collection of fixed anchors, Topos Pro world model scale-up, high-prevalence simulations of scale-based interactive output has taken shape.
It goes back not only to a simulation centre in Hangzhou, but also to a computer engine that promotes smart farewell “workshops” and moves towards generalization.
Original Ray Peak article, unauthorized reproduction. For further information, see the information below.
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.