World Labs Trained Zero-Data Robot Policies That Ran an Hour on Hardware
Robots trained entirely inside a digital simulation — never touching a physical arm, never resetting a cable, never recovering from a dropped test tube — ran continuously on real hardware for an hour without human intervention, according to a World Labs R2S2R technical post the company published today, July 28, 2026. That headline result discloses the early performance data from World Labs’ Real-to-Sim-to-Real (R2S2R) framework — the first concrete benchmark from the spatial-intelligence company since its $1 billion Series B raised in February 2026 at a $5 billion valuation, bringing total capital raised to $1.23 billion, per its February 2026 funding announcement.
The paper matters because it directly challenges the way robotics teams currently budget their time: physical trials, hardware resets, and real-world demonstrations are the most expensive line items in robot development, and the consensus assumption has been that simulation can supplement but not replace them. World Labs’ results, if they hold at scale, suggest the assumption needs updating.
The robotics research community has spent years trying to optimize sim-to-real transfer — the step where a policy trained in simulation is moved onto physical hardware. The dominant approach has been domain randomization: vary the simulation’s lighting, texture, mass, and friction parameters until the policy learns to generalize across that variation, with the assumption that real-world conditions fall somewhere inside the randomized range. Tobin (2017) and Peng (2018) established this approach as the field’s standard, and it has since become foundational to how most robotics teams design their training pipelines.
The Problem the Field Has Been Solving Wrong
World Labs argues this framing starts in the wrong place. If the simulation itself is a poor model of physical reality — if it assumes rigid bodies when real cables and elastic materials deform, if it models contact forces inaccurately — then no amount of domain randomization compensates for the upstream failure. A 2026 Annual Reviews survey on robotics confirms this framing: most simulators “assume perfectly rigid bodies and joints,” making deformable object tasks — cables, packaging, elastic connectors — the hardest category for conventional sim-to-real transfer.
The solution, according to World Labs’ R2S2R framework, is to solve the upstream problem first: make the simulation accurate enough that the transfer question nearly answers itself.
How R2S2R Works: Two Directions, One Loop
The framework runs in two phases that form a closed loop.
The first phase — Real-to-Sim, or R2S — starts with a physical robot task. Engineers capture the robot, its sensors, the surrounding environment, the objects it needs to manipulate, and a set of task demonstrations. A generative world modeling system then reconstructs all of that as an interactive simulation. Critically, the reconstruction targets both visual fidelity and physical fidelity: not just how the scene looks, but how objects deform, resist, and respond when the robot makes contact with them. World Labs validates each reconstruction by running an identical open-loop action sequence in both simulation and reality side by side, comparing observations, object responses, and task outcomes.
The second phase — Sim-to-Real, or S2R — uses that physics-aligned simulation as a training and evaluation environment. Because the simulation is accurate, it can systematically vary what a physical lab cannot easily vary: object configurations, robot states, camera viewpoints, lighting conditions, physical material properties, and difficulty levels. It can also provide ground-truth supervision — precise object positions, contact forces, interaction outcomes under alternative actions — that is difficult or impossible to obtain reliably on hardware. One reconstructed task can serve multiple policy architectures and multiple robot platforms, turning each simulation into reusable infrastructure rather than a one-off experiment.
This system is described as policy-agnostic and embodiment-agnostic — it does not require a specific robot or a specific learning algorithm.
What the Demonstrations Actually Showed
World Labs demonstrated R2S2R across five task categories, each chosen specifically because it involves the physical properties that conventional simulators struggle most to reproduce: deformable cables, elastic connectors, articulated boxes, test tubes with tight geometric tolerances, and thin objects extracted from dense clutter.
On an ALOHA bimanual robot, the system trained a policy for bimanual box packing — folding and sealing cardboard boxes — entirely in simulation, with zero real-world training data, and then transferred it directly to the physical robot. The policy also ran robustly under active lighting perturbations — a direct test of generalization beyond training conditions.
Four tasks demonstrated fully autonomous one-hour operation on physical hardware with no human intervention:
A YAM arm performed cable manipulation — sliding, routing, plugging, and unplugging deformable cables through sustained contact. An RB-Y1 bimanual platform routed a power cord around a refrigerator, a contact-rich task requiring coordinated two-arm manipulation of a deformable object. A Flexiv arm transferred test tubes between holders under tight geometric tolerances. An xArm extracted markers and pencils one at a time from dense clutter — the singulation problem, where thin similar objects must be differentiated by touch and position.
All five tasks trained policies with zero real-world data.
Simulation as a Screening Layer, Not Just a Training Tool
Beyond training, World Labs demonstrated R2S2R as a high-throughput evaluation engine — and this may be the more immediately practical result for robotics teams that already have policies they want to evaluate.
The team ran four policy architectures — Diffusion Policy (DP), ACT, π0, and π0.5 — on a bimanual cube-handover task using the ALOHA robot platform. Each architecture was evaluated across multiple checkpoints under both in-distribution (ID) and out-of-distribution (OOD) conditions. Simulated evaluation used 2,000 trials per checkpoint; real-world evaluation used 100.
The result was strong rank-order correlation: policies that performed better in simulation also performed better on hardware; checkpoint rankings were preserved across the two environments; and failure regions in simulation matched failure regions in reality. A simulation does not need to exactly match real-world success rates to be useful — it needs to preserve the relative performance ranking of the policies being compared and identify where each one fails. R2S2R appears to satisfy that weaker but practically sufficient condition.
The practical implication for robotics teams: rather than burning hardware time to evaluate every training checkpoint, teams can screen the vast majority of candidates in simulation and reserve physical testing for the most promising policies. The paper positions this as a means to turn robot development iteration from hardware-bound to simulation-first.
Why This Is Bigger Than One Paper
The economic argument embedded in R2S2R is worth stating explicitly, because the draft buries it: a simulation reconstructed once can serve many policies and many robot embodiments. Each iteration — new model, new hardware platform, new evaluation run — uses the same underlying world representation, with the per-iteration cost falling sharply compared to physical trial-and-error.
This mirrors the structural dynamic that defined large language model development. LLMs achieved their capabilities partly because internet-scale text — collected once, at enormous cost — could be amortized across all downstream tasks. R2S2R proposes an analogous amortization for robot tasks: invest in building a physics-accurate real-to-sim representation of a task once, then run thousands of training and evaluation iterations inside it. The robotics teams that invest earliest in high-fidelity real-to-sim reconstruction infrastructure are building an asset that compounds over time — a structural, repeatable cost advantage over teams that remain dependent on hardware data collection.
World Labs co-founder Fei-Fei Li, who created the ImageNet dataset that sparked the deep learning era, has framed this as the same problem class at a different layer: data, not algorithms, has historically been the binding constraint. In robotics, simulation alignment is becoming what dataset curation was for vision — the infrastructure layer whose quality determines everything built on top of it, as the company argued in its June 2026 taxonomy of world models.
Where World Labs Sits in a Crowded Field
The robotics simulation space is contested. NVIDIA’s Cosmos 3 world foundation model — now at its third iteration — generates synthetic robot training data from video and physics models, with FieldAI and Skild AI among the companies using it in production, according to the NVIDIA Physical AI announcement. Physical Intelligence’s π0 and π0.5 policy architectures, which World Labs evaluated as comparison baselines in its own benchmark, represent a different axis of the problem: policy architecture and open-world generalization, rather than simulation fidelity, as described in the Physical Intelligence π0.5 model paper.
The key distinction in World Labs’ positioning is that Cosmos generates synthetic data from generative video models, while R2S2R reconstructs real physical tasks as physics-accurate simulations. These approaches are not mutually exclusive: the NVIDIA Global Robotics announcement confirmed that World Labs is using Isaac Sim to validate its generative world models alongside its own R2S2R pipeline, suggesting the two stacks are more complementary in practice than the competitive framing implies.
The timeline is also accelerating from the commercial side. BYD confirmed on July 28, 2026 — the same day World Labs published its R2S2R results — that it will debut its first humanoid robot at its Di Space experience venues in early August 2026. Nearly 20 major automakers have now entered the humanoid robot race, with 2026 described as the critical window for transitioning from proof-of-concept to scaled deployment. The commercial pressure for simulation-trained policies to reach production-grade reliability is real, and it is building faster than the research calendar.
What R2S2R Does Not Yet Show
The published results are internally generated by World Labs and have not been independently replicated or peer-reviewed as of today’s publication. The one-hour autonomous operation benchmark is specific to pre-reconstructed tasks — environments that were deliberately captured and calibrated before training. The paper does not demonstrate R2S2R generalizing autonomously to novel environments that were not pre-reconstructed, which is the harder generalization challenge for production deployment.
The paper also acknowledges the key open research question: how does world model performance scale with additional interaction data and compute? The LLM era was defined by scaling laws — measurable relationships between compute investment and capability improvement. World Labs flagged studying the analogous scaling question for physical simulation as an important next direction, and the answer will determine how much of the pipeline’s current cost structure persists at scale.
What Robotics Teams Can Do With This Information
For a team building robot policies today, R2S2R’s most immediately actionable implication is the evaluation result. Even before committing to simulation-only training, teams can evaluate whether a physics-accurate simulation of their specific task can reliably predict which checkpoints will perform best on hardware. If the rank-order correlation holds for their tasks, simulation-first evaluation immediately reduces hardware time — a concrete, measurable payoff that does not require waiting for zero-data training to mature.
World Labs described the system as policy-agnostic and embodiment-agnostic, which means the same simulation infrastructure can in principle serve teams running different policy architectures — DP, ACT, π0, π0.5 — without rebuilding the world for each one. The company is actively hiring researchers and engineers and has indicated plans to extend R2S2R to more diverse environments and increasingly complex manipulation tasks, as noted in its R2S2R technical post. Interested researchers can view open positions on the World Labs careers page.
Frequently Asked Questions
What does “zero real-world training data” actually mean for the World Labs robots?
It means the neural network policies that controlled the physical robots were trained exclusively on data generated inside a simulation — no physical robot arm performed a task during the training process. The simulation itself was built by scanning and reconstructing a real physical task (capturing the robot, environment, and objects), but the policy training that followed used only simulated trials, not hardware trials. The robots were then deployed directly to physical hardware without any fine-tuning on real demonstrations.
How is World Labs’ R2S2R different from what Nvidia’s Cosmos is doing?
The approaches operate at different layers of the same problem. Nvidia Cosmos generates synthetic robot training data from video foundation models — it can hallucinate plausible robot interactions from video prompts without requiring a physical reconstruction of a specific task. World Labs’ R2S2R starts by reconstructing a real physical task environment into a physics-accurate simulation, with the goal of matching not just visual appearance but actual physical dynamics (how cables deform, how objects slide, how contact forces transmit). The two approaches are not mutually exclusive: World Labs has confirmed it uses Nvidia’s Isaac Sim to validate parts of its own pipeline, and future systems will likely combine both.
What is the “real-to-sim gap” and why does it matter more than the sim-to-real gap?
The sim-to-real gap is the performance drop that occurs when a policy trained in simulation is deployed on physical hardware. The robotics community has focused on closing this gap with domain randomization — training policies across many simulated variations so the policy generalizes. World Labs argues the real-to-sim gap is the upstream problem: if the simulation doesn’t accurately model how real objects deform, how friction behaves, and how contact forces transfer, no amount of domain randomization can compensate. A simulation that is wrong in its physics produces policies that fail on hardware not because of the sim-to-real transfer, but because the simulation was never an accurate proxy for reality in the first place.
Will robotics teams be able to use R2S2R on their own hardware and tasks?
World Labs has not announced a public SDK, API, or open-source release of R2S2R as of July 28, 2026. The results published today are from an internal proprietary system. The company is actively hiring for its robotics and simulation teams, and its blog post states plans to extend R2S2R to more diverse environments and tasks. How the technology will be made available to external teams — commercially, through partnerships, or as an open-source framework — has not been disclosed. Interested engineers and researchers can monitor open positions on the World Labs careers page.
