Atlas turns spatial capture into a robot test loop

Atlas reconstructs a photographed real space and can generate RGB views and depth information for a robot moving along different routes, turning spatial capture into the start of a robot test loop.

According to woshipm, the system takes a camera position and movement trajectory in 3D space, then builds a scene from one or more images as depth maps. It can also produce point clouds or 3D Gaussian Splat. That makes its value less about producing a convincing virtual room than about creating a controllable setting for testing a proposed action before hardware attempts it.

The boundary matters. woshipm characterizes Atlas as reconstruction and spatial rendering moving toward simulation, not evidence of complete physical-world mastery. Early access was limited to some partners, and World Labs had not publicly released model weights or training data. It had also not publicly released pricing or a formal API for general developers. The practical question is therefore narrower: when can a reconstructed world predict an action well enough to improve the next real attempt?

Atlas can reconstruct a scene before it can predict a task

Atlas begins with scene state rather than a finished camera clip. It accepts a camera position and movement trajectory in 3D space, according to woshipm, and can reconstruct spaces from one or more images. It can output depth maps and point clouds. It also produces 3D Gaussian Splat. In a robot demonstration, woshipm says Atlas reconstructed an ordinary-camera view of a real environment, then produced possible RGB and depth views along alternative robot routes.

A convincing view does not establish that the underlying state predicts an action. That gap matters when a rendered route becomes a hardware decision.

ifanr reports that World Labs claims video generation up to 1 minute at 1440p from 1 to 6 ordinary photos and a specified camera trajectory. The same account says inferred unseen areas depend on model priors. Larger viewpoint changes and longer paths can produce geometric drift or altered object shapes. Textures can flicker too.

World Labs also presents Atlas as a Real-to-Sim system that can form a high-precision 3D physical space from two 24-frame videos. It claims RGB and depth views for an onboard robot camera, plus force feedback for an arm touching a rigid box or pulling a hinged cabinet door. But, as woshipm notes, early access was limited at release. Some partners received it. There was no public full paper or model weights. Training data were not public. Pricing was undisclosed, and no formal API was available. Those limits leave reconstruction and rendering claims short of public proof of complete physical simulation.

How four world-model approaches connect simulation to real-world work

Small worlds give a robot a testable boundary

A robot test loop needs a boundary. It needs scale later. Ren Yunfan argues in the outlet that world models should start with limited, agent-relevant "small worlds," rather than attempt to understand everything. For a team, that means defining a closed real-world scenario and the task the agent must handle inside it. WorldMind plans to work with industry partners on one or two such scenarios, linking data collection to training and deployment. Feedback would follow.

The boundary makes failure interpretable.

WorldMind separates its model into knowledge and memory. It also separates inference from calibration. Its memory-enhanced model ranked first on WorldArena1.0's open-source leaderboard and stayed in the top three on multiple WorldArena2.0 tracks, according to the outlet. Those results matter as evidence of model capability, but they do not establish reliable transfer for every physical task. A task-specific environment gives calibration a practical target: the conditions under which an action is expected to work.

Earlier world-model systems also show why a manageable state matters. The system described by woshipm compressed game frames into 32-dimensional latent vectors, then used a recurrent network to predict later vectors. DreamerV3 used fixed hyperparameters across more than 150 tasks. Broad task coverage is useful, yet it does not remove the need to specify which situation a robot must model accurately enough to act.

General physical-AI models expand the available inputs. Nvidia's Cosmos 3 supports text, images, video, ambient sound and actions, while the 4 billion-parameter Cosmos 3 Edge targets local robotic world simulation. The practitioner's question remains narrower: can this defined environment predict the particular action that matters?

The useful architecture separates planning from execution

A world model is best treated as an internal representation that predicts how an environment changes after an action, according to woshipm. That representation need not come from one model type. Latent-state systems can compress relevant conditions. Video-predictive systems can forecast visible consequences. Spatial systems can work directly with 3D space, sensor inputs, and robot actions. Each route can inform planning, but none automatically supplies dependable execution.

The architectural split is simple: maintain state and choose a goal-directed action, then hand the action to a specialist that can perform it.

woshipm describes World Labs' distinction among renderers, simulators, and planners. Renderers produce observations. Simulators preserve world state. Planners select actions around goals. For a robot task, that separation prevents a convincing generated view from being mistaken for a reliable control policy. A planner can use latent predictions to consider consequences. It can also use video or spatial predictions, while an execution model focuses on the physical details that make one action succeed.

Future Far Away's design illustrates the practical value of that division. Its AVLA model breaks complex work into hundreds of shorter expert models, including grasping and fine manipulation. Its Self-Evolving WAM uses predicted outcomes as training data. It also uses actual results and failure cases from household tasks. The action layer therefore does not remain fixed after planning; it is revised against what the hardware actually did.

That loop has operational stakes. Reports indicate robots in about 500 Shanghai households, while Zhang Yi claims maintenance-free operation moved from one day to one to two months. The relevant measure is not a broad claim of world understanding. It is whether planning and specialized execution keep improving the next physical attempt.

A sim-to-real threshold must be defined per action

The threshold is not a single score for a world model. It is a repeatable transfer test for a bounded action: navigate a known route or reach a cabinet handle. For visual navigation, compare RGB and depth views predicted for an onboard camera with what hardware sees along the route; World Labs makes that Atlas rendering claim, according to ifanr. The ifanr author argues that larger viewpoint changes and longer generation paths increase local geometric drift and texture flickering. Completion on the real route is the measure.

That tests views, not contact.

Force feedback needs a stricter gate. World Labs claims Atlas can simulate force feedback when a robotic arm touches a rigid box or pulls a hinged cabinet door, ifanr reports. Each contact task should be run on hardware after simulation, with success defined before testing. If force or friction determines the result, convincing images cannot substitute for a real trial. The same applies when deformation or recovery determines it.

Variation should remain tied to the action boundary. The outlet says KnowinDream expands limited interaction experience into training distributions covering different spaces and lighting conditions. It also covers materials and task combinations. An evaluator can vary those conditions around the same action, then check whether it transfers. The outlet describes KNOWIN-X1 with prone operation on floors and standing operation at cabinets. Its arms each have 7 degrees of freedom. Its maximum dual-arm payload is 6 kg. Transfer is demonstrated only for the action and hardware configuration exercised, within the stated variation range.

Choose a world-model workflow by its ability to close the real-world loop

• You need to explore many variations of a bounded physical task-different spaces, lighting, materials, and task combinations-but have limited real-world interaction data. Use a task-specific data-expansion workflow, not a claim of general understanding. Nuoyin Intelligent describes KnowinDream as expanding limited real-world experience into distributions covering those variations, within a five-module chain from data generation to real-world feedback.
• Your main requirement is reconstructing a site and testing camera or robot routes before deployment. Prioritize controllable spatial inputs and outputs: camera poses or trajectories in, and RGB, depth, point clouds, or 3D Gaussian Splat outputs. Atlas is presented as accepting camera trajectories and reconstructing spaces from images, but its availability was limited to selected Early Access partners and the sources do not establish general real-world predictive reliability.
• A robot must improve after failures in homes or another closed operating environment. Choose a system designed to capture predictions, actual results, and failure cases, then retrain and redeploy. Future Far Away's Self-Evolving WAM uses all three as household-task training data; WorldMind similarly plans a data-collection, training, deployment, and feedback loop in one or two closed real-world scenarios.
• You are evaluating a visually impressive generated scene or video as evidence that a model can guide hardware. Do not treat rendering quality as sufficient. World Labs' own taxonomy separates renderers, which generate observations, from simulators, which maintain world states, and planners, which select goal-directed actions. The source material also notes that larger viewpoint changes and longer generation paths can increase geometric drift, object-shape changes, and texture flickering in Atlas outputs.

Closed loops need records, safeguards, and an exit to reality

A closed loop needs a record of what the system predicted, what happened on hardware, and which correction changed the next attempt. Without that history, a failure can be absorbed into a growing training set without showing whether the simulator or planner was wrong. It can also show whether the policy failed in execution. Responsibility becomes harder to assign when a plausible prediction leads to an incorrect household action.

The record must also be editable and versioned.

A report describes WorldMind as seeking industry partners for a cycle that begins with collection. It then moves through training before deployment and feedback in one or two closed real-world scenarios. Nuoyin Intelligent's GLOW system is presented as five modules connected in continuous iteration, from generated data through real-world feedback. Its KnowinDream component broadens limited interaction data across spaces and lighting conditions. It also covers materials and task combinations, while other components cover planning and recovery. Other components execute actions.

That expansion creates an operational burden. Each synthetic variation needs a link back to the conditions it represents, and each real failure needs enough labeling to distinguish a bad prediction from a bad action. Safeguards matter most where correction is expensive: private household captures, repeat visits, and actions around residents. A report says that Nuoyin Intelligent launched KNOWIN-X1 as a household robot, while Future Far Away's robots entered about 500 Shanghai households. Future Far Away uses predictions, actual results, and household failure cases as training data.

Vendors also need an exit to reality: a defined point where the system stops inferring and seeks confirmation, human intervention, or a new capture. Public scrutiny is difficult when key materials remain unavailable. According to woshipm, World Labs had not publicly released a full paper, weights, or training data at release. It also had not released pricing or a formal API for general developers. A controllable world model therefore requires audit trails around its corrections, plus clear limits on what a simulated success is allowed to authorize.

Set up the loop around a bounded household task, then keep the evidence from every miss. Future Far Away's Self-Evolving WAM uses predicted outcomes, actual results, and failure cases from household work as training data, according to reporting.

The useful test is not a convincing demonstration. Ask which actions are decomposed into specialist models: Future Far Away's AVLA separates complex work into hundreds of shorter expert models, including grasping and fine manipulation. Then track how long hardware runs before intervention. Zhang Yi says its robots moved from one day between on-site maintenance visits to one to two months. About 500 Shanghai households provide a route for collecting execution data; that feedback path matters more than a simulated scene alone.

For readers outside China

• Availability: Public availability outside China is not disclosed in sources. Atlas was available in Early Access only to some partners, with no formal API for general developers at the time described. Future Far Away's robots had entered about 500 households in Shanghai. KNOWIN-X1 was expected to launch formally in 2027.
• Pricing: Atlas pricing had not been publicly released at the time described. Future Far Away reported a robot cost of 20,000 to 30,000 yuan and household trials at 3,000 yuan. KNOWIN-X1 was described as expected to cost tens of thousands of yuan; its final price was not disclosed in sources.
• Closest Western equivalents: Nvidia Cosmos 3: a general-purpose physical-AI model that natively supports text, images, video, ambient sound, and actions; Cosmos 3 Edge is a 4 billion-parameter model for local robotic world simulation.; DreamerV3: a world-model reinforcement-learning system reported as using fixed hyperparameters across more than 150 tasks.; OpenAI Sora: a video-generation system, relevant as a contrast case rather than proof of a controllable simulator; OpenAI reportedly shut it down around March 24, 2026.
• Data residency: The source material does not cover data residency, hosting locations, cross-border data transfer, retention, or enterprise privacy controls for Atlas, WorldMind, Nuoyin Intelligent, or Future Far Away.

Sources

• 36kr 硬氪专访丨公司退市、裁员10万人，前掌门教育CEO创业家庭机器人，获10亿融资 https://36kr.com/p/3952867368648073
• woshipm 世界模型元年：一半火焰，一半冷水 https://woshipm.com/ai/6461298.html
• woshipm 李飞飞发布Atlas：世界模型，到底是个啥？ https://woshipm.com/ai/6459382.html
• 36kr 硬氪首发 | 99年同济副教授、前大疆骨干创业世界模型，成立不到一月融资数千万 https://36kr.com/p/3973553316081920
• 36kr 36氪首发 | 天使++轮融资5亿，前华为生成式大模型负责人把家务机器人折进40厘米 https://36kr.com/p/3933076448558212
• ifanr 刚刚，李飞飞掀桌！全球首个多模态世界模型发布，几张照片省掉几百个机位 https://ifanr.com/1678079