Paper deep dive

PlayWorld: Learning Robot World Models from Autonomous Play

Tenny Yin, Zhiting Mei, Zhonghe Zheng, Miyu Yamane, David Wang, Jade Sceats, Samuel M. Bateman, Lihan Zha, Apurva Badithela, Ola Shorinwa, Anirudha Majumdar

Year: 2026Venue: arXiv preprintArea: cs.ROType: PreprintEmbeddings: 79

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 3/13/2026, 12:57:46 AM

Summary

PlayWorld is a scalable framework for training high-fidelity, action-conditioned video world models for robotics. By utilizing autonomous robot self-play—guided by a vision-language model (VLM) and executed by a vision-language-action (VLA) policy—the system collects diverse, contact-rich interaction data without human supervision. This approach overcomes the limitations of success-biased human demonstration datasets, enabling improved physical consistency in video predictions, fine-grained failure prediction, and significant performance gains in downstream reinforcement learning and policy evaluation.

Entities (5)

PlayWorld · framework · 100%Vision-Language-Action Policy · component · 98%Vision-Language Model · component · 98%DROID · dataset · 95%Stable Video Diffusion · model-architecture · 95%

Relation Signals (4)

PlayWorld → improves → Reinforcement Learning

confidence 95% · PlayWorld enables reinforcement learning in the world model, improving policy performance by 65%

VLM → proposestasksfor → VLA

confidence 95% · the VLM and VLA iteratively propose and execute tasks

PlayWorld → trainsusing → Autonomous Robot Play Data

confidence 95% · PlayWorld, a simple and scalable world modeling framework built on autonomously gathered play data

PlayWorld → usesbackbone → Stable Video Diffusion

confidence 95% · we adopt a pre-trained stable video diffusion (SVD) backbone

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Action-conditioned video models offer a promising path to building general-purpose robot simulators that can improve directly from data. Yet, despite training on large-scale robot datasets, current state-of-the-art video models still struggle to predict physically consistent robot-object interactions that are crucial in robotic manipulation. To close this gap, we present PlayWorld, a simple, scalable, and fully autonomous pipeline for training high-fidelity video world simulators from interaction experience. In contrast to prior approaches that rely on success-biased human demonstrations, PlayWorld is the first system capable of learning entirely from unsupervised robot self-play, enabling naturally scalable data collection while capturing complex, long-tailed physical interactions essential for modeling realistic object dynamics. Experiments across diverse manipulation tasks show that PlayWorld generates high-quality, physically consistent predictions for contact-rich interactions that are not captured by world models trained on human-collected data. We further demonstrate the versatility of PlayWorld in enabling fine-grained failure prediction and policy evaluation, with up to 40% improvements over human-collected data. Finally, we demonstrate how PlayWorld enables reinforcement learning in the world model, improving policy performance by 65% in success rates when deployed in the real world.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

79,125 characters extracted from source content.

Expand or collapse full text

PlayWorld: Learning Robot World Models from Autonomous Play Tenny Yin 1 , Zhiting Mei 1 , Zhonghe Zheng 1 , Miyu Yamane 1 , David Wang 1 , Jade Sceats 1 , Samuel M. Bateman 1 , Lihan Zha 1 , Apurva Badithela 1 , Ola Shorinwa 1 , Anirudha Majumdar 1 1 Princeton University Action-conditioned video models offer a promising path to building general-purpose robot simulators that can improve directly from data. Yet, despite training on large-scale robot datasets, current state-of-the-art video models still struggle to predict physically consistent robot-object interactions that are crucial in robotic manipulation. To close this gap, we presentPlayWorld, a simple, scalable, and fully autonomous pipeline for training high-fidelity video world simulators from interaction experience. In contrast to prior approaches that rely on success-biased human demonstrations,PlayWorldis the first system capable of learning entirely from unsupervised robot self-play, enabling naturally scalable data collection while capturing complex, long-tailed physical interactions essential for modeling realistic object dynamics. Experiments across diverse manipulation tasks show thatPlayWorldgenerates high-quality, physically consistent predictions for contact-rich interactions that are not captured by world models trained on human-collected data. We further demonstrate the versatility ofPlayWorld in enabling fine-grained failure prediction and policy evaluation, with up to 40% improvements over human-collected data. Finally, we demonstrate howPlayWorldenables reinforcement learning in the world model, improving policy performance by 65% in success rates when deployed in the real world. Keywords: Video World Models, Autonomous Data Collection Website: robot-playworld.github.io 1Introduction Generative video models [1–4] hold tantalizing potential to serve as general-purpose data-driven simulators for robotics [5–7]. In principle, video models address some of the most critical challenges faced by traditional physics-based simulators by generating highly photorealistic outputs, simulating non-rigid-body interactions (e.g., deformable objects, liquids), providing a direct real-to-sim pipeline by conditioning video generation on images of real-world scenes, and affording the ability to close the sim-to-real gap by scaling data. An emerging line of work in robotics [8–14] seeks to reap these benefits by fine-tuning video generation backbones pre-trained on internet-scale data [1,15] with robotics datasets, and using the resulting models for scalable data generation [8, 16–18], policy evaluation [9, 11, 14, 19], and planning [10, 13, 20]. Despite this progress, state-of-the-art video models remain far from reliable world simulators. While capable of generating physically consistent, photorealistic, and long-horizon rollouts in non-interactive scenes (e.g., autonomous driving [21,22]), they remain vulnerable to hallucinations when simulating contact-rich inter- actions: from objects that duplicate when grasped, to ones that appear, disappear, or move and deform in unrealistic ways upon contact [3,14,23–29]. As a result, the use of video models in robotic manipulation has largely been restricted to tasks with minimal contact and focused around high-level language-following abilities [9, 14]. In this paper, we take a data-centric perspective on training high-quality video models for robotic manipulation. Our core hypothesis is that the aforementioned challenges with simulating contact-rich dynamics stem from the data on which video models are trained or fine-tuned. In particular, existing action-conditioned video models for manipulation (e.g., [9,11,19]) are almost exclusively trained on human demonstration datasets [30,31] developed for imitation learning. This inherently limits data coverage to a narrow distribution of states focused along successful task executions. As a result, this data provides video models with little supervision on complex contact dynamics and state transitions that counterfactual actions might introduce [32,33]. Small 1 arXiv:2603.09030v2 [cs.RO] 11 Mar 2026 Figure 1 We introducePlayWorld, a scalable framework for training high-fidelity video world models from autonomous robot play that enables fine-grained dynamics prediction for accurate policy evaluation, and online reinforcement learning policy fine-tuning that yields strong real-world success rate improvements. prediction errors at critical contact events can thus compound quickly into qualitatively divergent policy rollouts which are often heavily biased towards successful executions seen in the data (Sec. 4). How can we obtain the most useful data for learning high-quality world models? From a data-centric standpoint, we argue that the right training corpus needs to (i) be generated through directed exploration that continually expands coverage of diverse contact events, and (i) support efficient scaling, so exploration can grow naturally with minimal manual engineering or supervision. Inspired by work in developmental psychology [34,35], we ask: can robots actively curate a diverse set of experiences for training high-quality video models through autonomous play [36–38]? By interacting with objects in a semi-structured fashion without human supervision, autonomous play data can (i) induce broad coverage of relevant state transitions that might occur during policy execution, and (i) be highly scalable in terms of both quantity and diversity of objects and scenes. These advantages make autonomous play a compelling data-collection paradigm for contact-rich world modeling. Contributions. We introducePlayWorld, a simple and scalable world modeling framework built on autonomously gathered play data — trajectories generated from following loosely specified instructions that result in semi-structured interactions with objects in the scene. At a high level,PlayWorlduses a vision- language model (VLM) to propose diverse task instructions based on the robot’s scene, which are subsequently executed using pre-trained language-conditioned policies. With this framework, we make the following contributions. (1) Through careful empirical investigation, we demonstrate thatPlayWorldproduces significantly more diverse contact events, object states, and failure modes compared to human-collected data. (2) By limiting human involvement in periodically monitoring or resetting the scene, we show highly scalable data collection including unsupervised nightly collection. (3) Through comprehensive evaluations on various manipulation tasks, we demonstrate thatPlayWorld can simulate realistic physical interactions with fine-grained predictive accuracy, resulting in significant improvement on multiple perceptual metrics compared to baselines. (4) We present a data scaling study that shows continued improvements in downstream visual accuracy metrics withPlayWorld, even at 5×the scale at which performance saturates using human demonstration data. (5) We demonstrate that video models trained usingPlayWorldoutperform competing baselines in policy evaluation and RL fine-tuning applications, offering up to 65% improvement in real-world success rates over the pre-trained policy. To our knowledge, PlayWorldrepresents the first work to demonstrate the viability and benefits of autonomous robot play data for training action-conditioned video models. 2 2Related Work 2.1World Models for Robotics World models [39–41] learn predictive models of environment dynamics, enabling agents to reason and plan beyond purely reactive policies [42–44]. A large body of work in model-based reinforcement learning (RL) has demonstrated that large-scale interaction data with broad state-action coverage is necessary for training effective latent-space world models, which is often collected via on-policy or curiosity-driven exploration in simulation [42,45–47]. However, achieving the same levels of scale via exploration in real-world environments is challenging, posing a major bottleneck to collecting diverse interaction data required for training generalizable models [48]. As a result, many robotic world models are trained from either manually collected on-policy roll-outs [10,18,49,50] or large offline datasets [9,14,19,51], which are challenging to scale on hardware and can be biased toward narrow behaviors, limiting their ability to model the dynamical effects of diverse actions [33]. We address this limitation by designing a system to collect large-scale exploratory interaction data on hardware, expanding coverage over states and actions for learning generalizable dynamics. 2.2Policy Evaluation and Improvement with Video Models A growing body of work explores video generation models as embodied world models, leveraging large-scale pretraining for rich visual priors [1,2] and pixel-space future prediction as a versatile interface across tasks, embodiments, and downstream objectives [6,7,20,52]. This makes video models attractive for policy evaluation and improvement via world model rollouts, potentially reducing reliance on expensive real-world interaction. Prior work has primarily leveraged action-conditioned video prediction in two ways: (i) generating synthetic trajectories to augment training or fine-tuning [8,9,18,53], and (i) using imagined rollouts to score or optimize policies without repeated hardware trials [9,11,14,19,54–57]. However, the utility of these approaches is ultimately limited by their prediction fidelity under the policy-induced distribution [58]. In contact-rich robotic manipulation and high-DOF control, even small modeling errors can quickly lead to hallucinations in closed-loop rollouts, thereby resulting in inconsistent policy rankings and unstable reinforcement learning updates [59]. We address these limitations by developing an end-to-end system for scalable data collection and model training that prioritizes broad coverage of contact-rich interaction events, building a large, diverse corpus that approximates the outcome distribution induced by arbitrary policies. 2.3Leveraging Play Data in Robotics Play data, which consists of unstructured, task-agnostic interactions with broad coverage of interaction modes beyond narrow expert demonstrations, has been explored in prior works to enable planning [36,60,61], representation learning [62–64], policy learning [37,65–69], and finetuning pretrained models [70,71]. Recent work has also leveraged play data at a smaller scale to train world models that support policy improvement [72]. These approaches typically require additional human-collected play demonstrations or policy rollouts that depend on manual supervision, which fundamentally limits scalability and the breadth of interaction diversity [36,37]. In contrast, our work uses fully autonomous play data collection with generalist robot policies, enabling continuous, large-scale acquisition of diverse contact-rich interactions without supervision, transforming play data into a scalable resource for training high-fidelity world models. Most related to our work, SOAR [73] uses VLM-proposed tasks to autonomously collect data for improving language-conditioned policies In contrast, we use the collected play to learn a world model, converting finite real experience into effectively unlimited imagined interactions that can more flexibly support downstream policy applications. 3PlayWorld: Training Video World Models with Robot Play We now introducePlayWorld, a framework for training action-conditioned video models that can predict diverse contact dynamics with fine-grained precision. In this section, we first formalize the world modeling problem as learning generalizable dynamics under distributional mismatch. Then, we introduce the design of our autonomous data collection system that can reliably collect hours of interaction data with no human supervision. Finally, we discuss model architecture and training designs that allow us to effectively learn from large quantities of heterogeneous interaction data. 3 Figure 2PlayWorldSystem Diagram. Left: Autonomous data-collection pipeline in which the VLM and VLA iteratively propose and execute tasks. Right: Video world-model backbone and the setup for policy evaluation and fine-tuning. 3.1Generalizable Dynamics Learning through Play Lets t ∈Sdenote the proprioceptive state of the robot,a t ∈Adenote actions, ando t = (o 1 t ,...,o K t )∈O denote image observations corresponding toKdifferent views (e.g., multiple overhead cameras and a wrist camera) at timet. We model the robot-environment dynamics using a stochastic action-conditioned video model: p θ (x t+1 | x 1:t ,a 1:t ),(1) wherex t = (s t ,o t ). As highlighted in Sec. 2, prior work trains such models on human demonstration data, thereby optimizing predictive accuracy only under the visitation distribution (x 1:t ,a 1:t ,x t+1 )∼ D exp induced by the expert. As a result, the learned dynamics can suffer from distributional bias when queried under new policies and collapse to predictions that are likely underD exp [32,74]. In contrast, our goal is to train a video model that remains accurate across a broad range of policies. In order to achieve this, we expand the visitation distribution of the training data through data collected under a language-conditioned policyπ play (a|x,ℓ), whereℓ ∈ Ldenotes a natural language instruction. By generating highly diverse language instructions in a broad range of initial conditions encountered during autonomous roll-outs, this policy yields a set: D play =(x t ,a t ,x t+1 ) a t ∼π(·|x t ,ℓ),ℓ∼p(L) (2) that can better approximate visitation distributions induced by a wide range of policies. 3.2Autonomous Robot Play Data Collection Next, we discuss a practical instantiation ofπ play that can enable reliable and continuous autonomous data collection on hardware. In order to obtain useful and scalable play data, we outline three key design requirements: 1.The robot needs to engage in diverse interaction with objects in the scene (in order to obtain meaningful coverage as data collection expands) 2. The system must reliably prevent and recover from potential failures when human supervision is unavailable 3. The system should be amenable to diverse objects and generalize to diverse language instructionsℓ without requiring manual engineering Building on these requirements, we develop a practical system atop the DROID manipulation setup [31] that requires no additional hardware modifications and supports up to eight hours of continuous, fully autonomous data collection. As illustrated in Fig. 2, our system is composed of two interacting components: a task 4 proposer that generates tasks to engage the robot in meaningful interactions with objects, and a task executer that performs the task to generate interaction data. Task Proposer. We instantiate the proposer using a vision–language model (VLM) [75] that generates diverse, scene-grounded instructions from the robot’s current image observation(s)o t . At the start of each episode, the VLM is prompted to produce a natural-language instructionℓ(e.g., “push the object forward” or “stack the object”). To promote data diversity while preserving executability, we prompt the VLM to add small perturbations to nominal commands (e.g., alternative verbs or object descriptions) to probe different behavior modes. Compared to novelty- or uncertainty-driven on-policy exploration, our language-conditioned play requires no reward design or auxiliary training, and generates semantically grounded diversity by varying task intent (rather than injecting action noise), yielding visitation distributions that better match test-time policy behavior. More details on implementation and example prompts can be found in Appendix A. Task Executer. We use a vision–language–action (VLA) policy [76] as the executer that can be conditioned on arbitrary natural-language instructions. While it succeeds on nominal commands (e.g., “pick up the object”), small perturbations in the instruction such as verb choice or unfamiliar object descriptions often lead to large variances in policy behavior [77]. We exploit this as a source of diversity: during long-horizon autonomous roll-outs, we intentionally randomize instructions under naturally-induced varied initial object poses, boosting the diversity of contact dynamics modes for interaction-rich data collection. Safety Filter and Resets. To enable reliable unsupervised execution for long periods of time, we utilize a lightweight safety filter that constrains the robot to operate within conservative workspace limits. Specifically, we enforce per-joint limits to prevent drastic motions, and prompt the VLM to detect when any object is drifting towards the boundary of the robot’s reachability limit, to instruct the VLA to “reset” the scene by retrieving that object back to the reachable workspace. Since the setup is task/object-agnostic, during play data collection involving large number of objects (such as in Sec. 4.6), we can simply add or remove arbitrary objects during collection. This simple design works well in practice and allows us to collect data for long durations including overnight. 3.3Model Architecture and Training Finally, we discuss our choice of world model architecture and a curriculum-based training scheme for learning from uncurated interaction data. Video Model Architecture. Given the datasetD play , our goal is to train a world model that produces high-fidelity visual outputs that simulate fine-grained interactions with the environment. Following [9], we adopt a pre-trained stable video diffusion (SVD) backbone [15] with factorized spatial and temporal attention, allowing us to inject fine-grained per-frame action conditioning to disentangle motion signals from frame appearances for strong controllability. We train our model to jointly predict three camera views to minimize the influence of partial observability, similar to [10,14], and initialize our model with weights pre-trained on the DROID dataset [31]. Then, we fine-tune the world model on the play datasetD play using the diffusion loss function: L ε = E x 0 ,t,ε h ∥ε−ε θ (x t ,t)∥ 2 i ,(3) whereε∼N (0,I). We train the fullPlayWorldmodel on 8x H200 GPUs with batch size of 64 for two days. Curriculum Learning. As we scale up play data collection to increase coverage of diverse states and transitions, the resulting dataset introduces two challenges: (1) high redundancy, where the majority of data are dominated by very similar transitions, and (2) multi-modality, where there also exists many rarely-occurring transitions that exhibits strong diversity (long-tail) [37]. This unbalanced and long-tailed distribution makes learning with standard training procedures difficult, causing models to easily overfit to simple patterns while failing to capture rare interactions [78, 79]. To enable a more balanced exposure, we adopt a curriculum learning setup [80,81] to feed training data into the model in order of (auto-rated) “difficulty”: initializing with frequently occurring free space motions and static contacts, and gradually biasing sampling towards rare, harder-to-learn interactions. Concretely, suppose we are given a small set of human-collected demonstrationsD exp =τ i N s i=1 , where each trajectory 5 Table 1 Per-category perceptual similarity metrics on interaction-centric benchmark.PlayWorldimproves prediction quality on contact-rich failure modes, with further gains from scaling and curriculum learning. Training MixSuccessMissed GraspSlideSlipDeformationCollision LPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPSSSIM Human Demo (6h)0.0840.8670.0800.8750.0900.8500.0900.8650.1080.8200.0860.852 Human Play (6h)0.0860.8610.0710.8690.0890.8640.0880.8670.1000.8310.0800.873 Robot Play (6h)0.0820.8700.0660.8830.0770.8650.0780.8710.0990.8310.0740.888 Robot Play (30h)0.0710.8730.0640.8870.0730.8760.0720.8790.0940.8330.0760.883 Robot Play (Curriculum)0.0700.8800.0660.8900.0710.8900.0700.8840.0930.8360.0720.893 Figure 3 Illustration of each test category from the interaction-centric benchmark in Table 1. τ =(x t ,a t ) T t=1 corresponds to successful task execution. LetD play = ̃τ j N p j=1 denote uncurated and unlabeled play data collected using the protocol from Sec. 3.2. Letf clip (·) denote a frozen CLIP [82] image encoder. For each observation o t , we compute an embedding z t = f clip (o t )∈ R d .(4) We first extract a set of representative success centroids fromD exp using all embeddings from success trajecto- ries,Z succ =z t | o t ∈D succ , and then applyK-means clustering to obtain centroidsC=c k K k=1 , c k ∈ R d . These centroids serve as prototypes of task-relevant transitions observed in successful executions. For each observation o in D play , we define its distance-to-success score as d(o) = min c k ∈C ∥f clip (o)− c k ∥ 2 .(5) Intuitively, smaller values ofd(o) indicate transitions that are visually and semantically closer to successful trajectories, while larger values correspond to less task-relevant or exploratory interactions. We then induce a curriculum by partitioning D play into R disjoint ranks based on this distance: D play = R [ r=1 D (r) , D (r) =o| d(o)∈ [δ r−1 ,δ r ),(6) whereδ r R r=0 are distance thresholds chosen such thatδ 0 = 0 andδ R = +∞. Lower ranks correspond to samples that are closer to success trajectories, while higher ranks contain increasingly out-of-distribution interactions. During training, we progressively expand the sampling distribution over ranks, starting from lower rand gradually incorporating higher-ranked samples. This curriculum mitigates redundancy in free-space motion and counteracts the long-tailed distribution of critical interaction dynamics, enabling the model to learn rare but essential transitions more effectively. 4Experiments In this section, we design experiments to explore the following questions: 1. DoesPlayWorldinduce more diverse object interactions compared to human-collected data (Sec. 4.2)? 2.CanPlayWorldimprove video prediction accuracy for diverse object interactions compared to models trained on human demonstration data (Sec. 4.3)? 3.CanPlayWorldenable fine-grained policy evaluation by reliably predicting outcomes across a broad range of policies and tasks (Sec. 4.4)? 6 4. Can PlayWorld enable policy fine-tuning through interactive roll-outs in the video model (Sec. 4.5)? 5.DoesPlayWorldresult in improved accuracy and generalization with data scale compared to human- collected data (Sec. 4.6)? 4.1Experiment Setup Evaluation Setup. To evaluate whetherPlayWorldcan produce more accurate dynamics predictions under realistic settings where object interactions are diverse and contact-rich, we carry out experiments on three distinct object sets featuring objects with different physical properties and interaction behaviors, each with a collection of possible tasks: • Set 1: Bowl, carrot, polar bear (put the carrot/polar bear into/out of the bowl) • Set 2: Rectangular block and cube (stack/unstack the block on top of the cube) • Set 3: Towel (fold/unfold towel) For our method (PlayWorld), we follow Sec. 3.2 to collect a total of 30 hours of task-agnostic autonomous robot play data containing diverse robot-object interactions combined across all 3 object sets. Baselines. We collect two types of teleoperation data as baselines: human demo data consisting of task-specific expert demonstrations, and human play data where the operator is instructed to freely interact with given objects in a task-agnostic manner. In total, we collect 6 hours of human demo data on the identical setup as robot play data. We initialize all models from the DROID-pretrained checkpoint in [9]. Our human demo baseline corresponds to the standard demonstration-only fine-tuning paradigm (as in [9,19]) designed in particular for our environment, enabling a controlled comparison that isolates the effect of interaction data. 4.2Does PlayWorld Induce Diverse Interactions? Figure 4 t-SNE analysis of training samples. Robot play data exhibits markedly broader behavioral coverage than human-collected trajectories. Colors indicate coarse inter- action modes assigned by a human annotator. First, we evaluate ifPlayWorldcan induce more di- verse interactions with objects compared to human- collected data, as outlined in Sec. 3.1. For this com- parison, we analyze a subset of autonomous play data alongside the human demo and human play baselines from object set 1 (Sec. 4.1) and provide a t-SNE visualization [83] on CLIP embeddings of the image observations. To better interpret the embed- ding space, we let a human annotator inspect the corresponding video snippet for each data point and assigns it to a coarse, human-interpretable inter- action mode when applicable. As shown in Fig. 4, autonomous play exhibits substantially broader be- havioral coverage, capturing diverse contact-rich events that are consistent with human annotations (e.g., missed grasps, collisions, and slips). In con- trast, human-collected data forms a more concen- trated cluster that largely reflects successful transi- tions. More results on conducting data clustering at larger scales can be found in Appendix B. 4.3Can PlayWorld Accurately Predict Contact Dynamics? Next, we evaluate if thePlayWorld’s broader data coverage enables training better video world models for predicting challenging object dynamics. To this end, we construct an interaction-centric benchmark with 500+ clips sampled from roll-outs generated by a diverse collection of 20+ robot policies, categorized into 6 behavior modes by human annotators (Fig. 3), containing both successful task executions and common failure modes that are critical in determining policy behavior. Given a short ground-truth observation–state 7 ! GT Baseline PlayWorld ! GT Baseline PlayWorld ! GT Baseline PlayWorld ! GT Baseline PlayWorld Real-World Hallucination Missed Grasp Real-World Hallucination Deformation Real-World Hallucination Collision Real-World Hallucination Slip . Figure 5 PlayWorld faithfully captures fine-grained dynamic behaviors.PlayWorld’s predictions closely match fine-grained dynamic interaction modes in the real-world, such as missed grasps, collisions, deformations, and slips. trajectory as conditioning input, we roll out the ground-truth actions in the world model and compare the generated vs. ground-truth future frames using visual metrics such as LPIPS [84] and SSIM [85] (results on other visual metrics can be found in Appendix C). As shown in Table 1, while the prediction quality for successful interactions is similar across different training mixtures,PlayWorldprovides consistent improvements for other dynamic interactions. In addition, scaling PlayWorld data from 6h to 30h further boosts performance. Notably, curriculum learning offered substantial advantage in improving prediction quality for more dynamic interactions, given the same training data. Fig. 5 provides some example predictions when rolling out a given action sequence from the ground-truth trajectory under different world models; we observe that predictions from the baseline methods often collapse to behaviors biased by the training data, with “hallucinated success” being the most common failure mode. 4.4Can PlayWorld Predict the Performance of Different Policies? Figure 6PlayWorld’s predicted policy outcome dis- tribution matches closely with real-world behavior. While high-level policy capabilities such as instruction following are relatively easy to assess, evaluating fine- grained visuomotor abilities is substantially harder: small errors in contact timing, object position, or contact forces can lead to qualitatively different outcomes during closed- loop execution. To rigorously test whetherPlayWorld can faithfully predict policy performance, we construct a diverse suite of 18 policies by (i) training diffusion policies (DPs) [86] from scratch and (i) fine-tuningπ 0 [87] using human demonstrations that vary in both quantity and quality. This produces policies spanning a wide range of success rates and exhibiting diverse failure modes (Fig. 6), providing a sensitive probe of whether a world model cap- tures subtle contact-rich dynamics under realistic settings. For each policy, we perform 20 real-world experiments and 50 simulated experiments in each world model. 8 Figure 7 Policy Evaluation Success Rate (SR) Correlation. Across diverse policies (architectures, training mixtures, and tasks),PlayWorld’s predicted success rates are most strongly correlated with observed success rates and show the lowest variability. Figure 8 Example Policy Interaction with different World Models. We provide closed-loop rollouts with two example policies to demonstrate how hallucinated success can impact prediction outcomes. As shown in Fig. 7,PlayWorldgeneralizes well to a broad family of policies with substantially different success rates and behaviors (despite training on roll-outs collected form a single autonomous play policy). Predicted success rates from PlayWorld are strongly correlated with real-world success rates (Pearson correlation: 0.8766). In contrast, baseline models are only capable of capturing a narrow set of failure modes. When policies induce interaction patterns that are underrepresented in their training distributions, baseline models tend to either regress toward familiar outcomes or generate unrealistic dynamics with noticeably degraded visual quality (Fig. 8), leading to large discrepancies between predicted and observed success rates. In particular, we find that human play data yields the worst performance and often produces blurry predictions, which suggests that out-of-distribution diversity that doesn’t match well with policy behavior could hurt performance. Beyond aggregate success rates, we further assess whetherPlayWorldcan produce meaningful predictions of both success and failure by comparing human-annotated failure modes for each set of experiments. In Fig. 6, we present real and predicted failure-mode distributions for several representative policies and find thatPlayWorld’s predicted distributions broadly align with empirical outcomes, whereas the behavior within the baseline models have large variances. 4.5Policy Fine-Tuning in the World Model Many reinforcement learning fine-tuning algorithms can substantially improve policy robustness, but are seldom deployed on physical robots because real-world interaction is expensive, safety-limited, and slow to 9 run. While many existing video world models are used to optimize policies by generating synthetic data for supervised fine-tuning [8,88,89], here we show that with high dynamics prediction quality,PlayWorld can effectively support in-model RL fine-tuning that confers substantial gains in real-world deployment with minimal sim-to-real gap, making real-world RL more practical and scalable. Figure 9 Fine-tuning Experiment Results.PlayWorld-finetuned policy demonstrates improved success rates and learns robust recovery behaviors beyond seen demonstrations. To support stable fine-tuning, we adopt Diffusion Steering via Reinforcement Learning (DSRL) from [90], which freezes the base diffusion policy and instead learns a lightweight latent-noise policyπ w that samples the initial diffusion noisew(re- placingw ∼N(0,I)) to steer the resulting actiona=π W dp (s,w) toward higher reward, while avoiding unstable backpropagation through the multi-step denoising chain. We train a simple progress-based reward function on small amounts of demonstra- tion data, and compute the difference in predicted progress at each frame as dense rewards. Despite the simplistic setup, we observe surprisingly effective fine-tuning results. Fig. 9 summarizes our results on performing fine-tuning over two example tasks. Starting from a base diffusion pol- icy trained with fewer than 10 demonstra- tions, performing online fine-tuning fully withinPlayWorldimproves real-world suc- cess rates by up to 65%, compared to fine- tuning in the baseline video model. In particular, we find that the fine-tuned model performs much more robustly on out-of-distribution initializations (including ones unseen during fine-tuning) and learns subtle recovery behaviors like scooping, as illustrated at the bottom of Fig. 9. In contrast, the baseline model trained on demonstrations offers less significant improvements and becomes unstable as the policy tries to “hack” the world model when failure modes are incorrectly identified, resulting in decrease in real-world success rates. We present more detailed findings in the Appendix D. 4.6Scaling and Generalization Finally, we evaluate whether PlayWorld delivers sustainable, scalable data collection for world-model learning by measuring continued performance gains as we increase data scale and object diversity. Figure 10 Scaling behavior for world models trained with different data mixtures. Data Scaling. While scaling up data might appear to be a straightforward recipe for training better models, we show that the effectiveness of scaling depends strongly on data coverage. As shown in Fig. 10, the video model trained withPlayWorldcontinues to improve on visual- quality metrics as we scale the dataset up to 30 hours. By contrast, baseline models show only marginal gains at comparable scale. This divergence highlights a central advantage of autonomous play: rather than repeatedly sampling narrow, success-focused trajectories, play data expands coverage over contact events, failure modes, and counterfactual transitions. As a result, additional data meaningfully broadens the learned dynamics model in- stead of reinforcing existing biases. These results suggest that PlayWorld supports effective scaling, where increased data directly translates into improved generalization. 10 ObjectsFruitBlocks LPIPS↓ SSIM↑ LPIPS↓ SSIM↑ 0%0.09910.83330.09780.8501 33%0.07940.86120.07130.8810 66%0.07610.86970.06840.8853 100% 0.07050.87570.06400.8948 Figure 11 Object Generalization Results. We observe strong improvements in prediction quality of interaction with unseen objects as we include more diverse items in the training data. Object Generalization. We next test whether increasing object diversity in the training corpus improves transfer to unseen objects. As shown in Fig. 11, as we increase the fraction of training objects from 0% to 100%, prediction quality improves consistently on held-out objects across both test categories, despite substantial shifts in appearance and geometry. This suggests thatPlayWorldlearns dynamics that transfer across object instances, rather than memorizing object-specific visual features. We attribute this generalization to PlayWorld’s ability to scale object diversity while still inducing rich contact interactions during play, providing supervision on shared interaction patterns (e.g., contact, slip, deformation) that recur across different objects. 5Conclusion We presentPlayWorld, an efficient, autonomous, and scalable pipeline for training high-fidelity action- conditioned video world models on interaction-rich robot play data. By collecting large volumes of diverse experience autonomously (including unattended overnight operation),PlayWorldyields broader coverage of contact events and a richer learning signal than success-biased demonstrations. The resulting video world model achieves strong dynamics fidelity on both nominal and off-nominal interactions, enabling reliable fine-grained policy evaluation and in-model RL fine-tuning. Finally, we observe continued gains with increased data scale and object diversity, suggesting a practical path toward high-quality robot simulators that support policy evaluation and improvement beyond hardware constraints. 6Limitations and Future Work Improving Data Collection. WhilePlayWorldcan generate many diverse interaction data at scale, the current collection strategy does not explicitly optimize for sample efficiency and can produce redundant trajectories. Future work could explore better active data collection strategies that prioritize high uncertainty or underrepresented interactions that can further reduce hallucinations and improve physical consistency. Additionally, the current system relies on the availability of a sufficiently capable play policy; designing specialized or adaptive play policies that explicitly target rare contact events could further enhance coverage and efficiency. Dynamic Discrepancies. Despite significant improvements, PlayWorld does not eliminate hallucinations entirely. Prediction errors can still arise from open-loop rollout discrepancies, fixed action horizons, and mismatches between the control modes used during data collection and policy evaluation, which can accumulate over long horizons and degrade closed-loop stability. Future work could investigate more flexible prediction horizons, hierarchical or receding-horizon rollouts, and tighter integration between action conditioning and control representations to better align imagined dynamics with real-world execution. Scaling Up. While we demonstrate strong scaling trends within our experimental setup, extending PlayWorld beyond a controlled lab environment remains an open challenge. Scaling to more diverse objects, scenes, robot embodiments, and real-world settings will require improved strategies for learning from increasingly heterogeneous data distributions. As data scale increases, improved curriculum design and data balancing 11 mechanisms will become even more important; developing principled methods for automatically defining curriculum that optimize both prediction fidelity and robustness to long-tail interactions remains a key direction for future work. Efficiently leveraging large-scale, multi-robot datasets and potentially incorporating non-robot video data for broader physical priors also presents exciting opportunities. Overall, we believe that PlayWorld showcases the promise of scaling autonomous play data for training high-fidelity world models that can serve as general-purpose simulators for policy evaluation and reinforcement learning. Acknowledgments Apurva Badithela is supported by the Presidential Postdoctoral Research Fellowship at Princeton University. The authors were partially supported by Apple Inc. and the NSF CAREER Award #2044149. Any views, opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and should not be interpreted as reflecting the views, policies or position, either expressed or implied, of Apple Inc. 12 References [1]Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 1, 3 [2]Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 3 [3]DeepMind. Veo-3: A text-to-video generation system with audio. Technical Report Tech Report, DeepMind / Google, 2025. 1 [4]Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1 [5] Siqiao Huang, Jialong Wu, Qixing Zhou, Shangchen Miao, and Mingsheng Long. Vid2world: Crafting video diffusion models to interactive world models, 2025. 1 [6] Google DeepMind. Genie 3: A New Frontier for World Models. Google DeepMind Blog, aug 2025. 3 [7]Zhiting Mei, Tenny Yin, Ola Shorinwa, Apurva Badithela, Zhonghe Zheng, Joseph Bruno, Madison Bland, Lihan Zha, Asher Hancock, Jaime Fern ́andez Fisac, et al. Video generation models in robotics-applications, research challenges, future directions. arXiv preprint arXiv:2601.07823, 2026. 1, 3 [8]Joel Jang, Seonghyeon Ye, Zongyu Lin, Jiannan Xiang, Johan Bjorck, Yu Fang, Fengyuan Hu, Spencer Huang, Kaushil Kundalia, Yen-Chen Lin, et al. Dreamgen: Unlocking generalization in robot learning through video world models. arXiv preprint arXiv:2505.12705, 2025. 1, 3, 10 [9]Yanjiang Guo, Lucy Xiaoyang Shi, Jianyu Chen, and Chelsea Finn. Ctrl-world: A controllable generative world model for robot manipulation. arXiv preprint arXiv:2510.10125, 2025. 1, 3, 5, 7, 19 [10]Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin, Yen-Chen Lin, Yunhao Ge, Grace Lam, Percy Liang, Shuran Song, Ming-Yu Liu, Chelsea Finn, et al. Cosmos policy: Fine-tuning video models for visuomotor control and planning. arXiv preprint arXiv:2601.16163, 2026. 1, 3, 5 [11]Yaxuan Li, Yichen Zhu, Junjie Wen, Chaomin Shen, and Yi Xu. Worldeval: World model as real-world robot policies evaluator. arXiv preprint arXiv:2505.19017, 2025. 1, 3 [12]Yao Feng, Hengkai Tan, Xinyi Mao, Chendong Xiang, Guodong Liu, Shuhe Huang, Hang Su, and Jun Zhu. Vidar: Embodied video diffusion model for generalist manipulation. arXiv preprint arXiv:2507.12898, 2025. [13] Junbang Liang, Pavel Tokmakov, Ruoshi Liu, Sruthi Sudhakar, Paarth Shah, Rares Ambrus, and Carl Vondrick. Video generators are robot policies. arXiv preprint arXiv:2508.00795, 2025. 1 [14] Gemini Robotics Team, Krzysztof Choromanski, Coline Devin, Yilun Du, Debidatta Dwibedi, Ruiqi Gao, Abhishek Jindal, Thomas Kipf, Sean Kirmani, Isabel Leal, Fangchen Liu, Anirudha Majumdar, Andrew Marmon, Carolina Parada, Yulia Rubanova, Dhruv Shah, Vikas Sindhwani, Jie Tan, Fei Xia, Ted Xiao, Sherry Yang, Wenhao Yu, and Allan Zhou. Evaluating Gemini robotics policies in a Veo world simulator, 2025. 1, 3, 5 [15] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 5 [16]Junbang Liang, Ruoshi Liu, Ege Ozguroglu, Sruthi Sudhakar, Achal Dave, Pavel Tokmakov, Shuran Song, and Carl Vondrick. Dreamitate: Real-world visuomotor policy learning via video generation. arXiv preprint arXiv:2406.16862, 2024. 1 [17]Shivansh Patel, Shraddhaa Mohan, Hanlin Mai, Unnat Jain, Svetlana Lazebnik, and Yunzhu Li. Robotic manipulation by imitating generated videos without physical demonstrations. arXiv preprint arXiv:2507.00990, 2025. [18]Yanjiang Guo, Tony Lee, Lucy Xiaoyang Shi, Jianyu Chen, Percy Liang, and Chelsea Finn. Vlaw: Iterative co-improvement of vision-language-action policy and world model, 2026. 1, 3 13 [19]Julian Quevedo, Ansh Kumar Sharma, Yixiang Sun, Varad Suryavanshi, Percy Liang, and Sherry Yang. Worldgym: World model as an environment for policy evaluation, 2025. 1, 3, 7 [20]Boyuan Chen, Tianyuan Zhang, Haoran Geng, Kiwhan Song, William T. Freeman, Jitendra Malik, Russ Tedrake, Vincent Sitzmann, and Yilun Du. Large video planner enables generalizable robot control, 2025. 1, 3 [21]Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving, 2023. 1 [22]Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas M ̈uller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1 [23]Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In International conference on machine learning, pages 1174–1183. PMLR, 2018. 1 [24] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360–378. Springer, 2024. [25]Rick Akkerman, Haiwen Feng, Michael J Black, Dimitrios Tzionas, and Victoria Fern ́andez Abrevaya. Interdyn: Controllable interactive dynamics with video diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 12467–12479, 2025. [26]Bo Ai, Stephen Tian, Haochen Shi, Yixuan Wang, Tobias Pfaff, Cheston Tan, Henrik I Christensen, Hao Su, Jiajun Wu, and Yunzhu Li. A review of learning-based dynamics models for robotic manipulation. Science Robotics, 10(106):eadt1497, 2025. [27]Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. [28] Zhiting Mei, Ola Shorinwa, and Anirudha Majumdar. How confident are video models? Empowering video models to express their uncertainty. arXiv preprint arXiv:2510.02571, 2025. [29]Zhiting Mei, Tenny Yin, Micah Baker, Ola Shorinwa, and Anirudha Majumdar. World models that know when they don’t know: Controllable video generation with calibrated uncertainty. arXiv preprint arXiv:2512.05927, 2025. 1 [30]Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 1 [31] Alexander Khazatsky et al. Droid: A large-scale in-the-wild robot manipulation dataset, 2025. 1, 4, 5 [32] Chaochao Lu, Biwei Huang, Ke Wang, Jos ́e Miguel Hern ́andez-Lobato, Kun Zhang, and Bernhard Sch ̈olkopf. Sample-efficient reinforcement learning via counterfactual-based data augmentation. arXiv preprint arXiv:2012.09092, 2020. 1, 4 [33]Silviu Pitis, Elliot Creager, and Animesh Garg. Counterfactual data augmentation using locally factored dynamics. Advances in Neural Information Processing Systems, 33:3976–3990, 2020. 1, 3 [34]Justine E. Hoch, Sinclaire M. O’Grady, and Karen E. Adolph. It’s the journey, not the destination: Locomotor exploration in infants. Developmental Science, 22(2):e12740, March 2019. doi: 10.1111/desc.12740. Epub 2018 Oct 8. 2 [35]Deena Skolnick Weisberg, Kathy Hirsh-Pasek, Roberta Michnick Golinkoff, Audrey K. Kittredge, and David Klahr. Guided play: Principles and practices. Current Directions in Psychological Science, 25(3):177–182, 2016. doi: 10.1177/0963721416645512. 2 [36] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play, 2019. 2, 3 [37] Zichen Jeff Cui, Yibin Wang, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. From play to policy: Conditional behavior generation from uncurated robot data, 2022. 3, 5 14 [38]Zhiyuan Zhou, Pranav Atreya, Abraham Lee, Homer Walke, Oier Mees, and Sergey Levine. Autonomous improvement of instruction following skills via foundation models, 2024. 2 [39] David Ha and J ̈urgen Schmidhuber. World models. 2018. doi: 10.5281/ZENODO.1207631. 3 [40] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination, 2020. [41]Nicklas Hansen, Hao Su, and Xiaolong Wang. Td-mpc2: Scalable, robust world models for continuous control, 2024. 3 [42]Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 3 [43]Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. Dino-wm: World models on pre-trained visual features enable zero-shot planning, 2025. [44] Kensuke Nakamura, Lasse Peters, and Andrea Bajcsy. Generalizing safety beyond collision-avoidance via latent- space reachability analysis, 2025. 3 [45]Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International conference on machine learning, pages 2778–2787. PMLR, 2017. 3 [46] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International conference on machine learning, pages 8583–8592. PMLR, 2020. [47]Nicklas Hansen, Hao Su, and Xiaolong Wang. Td-mpc2: Scalable, robust world models for continuous control. arXiv preprint arXiv:2310.16828, 2023. 3 [48] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In Conference on robot learning, pages 2226–2240. PMLR, 2023. 3 [49] Zhennan Jiang, Kai Liu, Yuxin Qin, Shuai Tian, Yupeng Zheng, Mingcai Zhou, Chao Yu, Haoran Li, and Dongbin Zhao. World4rl: Diffusion world models for policy refinement with reinforcement learning for robotic manipulation. arXiv preprint arXiv:2509.19080, 2025. 3 [50]Fangqi Zhu, Zhengyang Yan, Zicong Hong, Quanxin Shou, Xiao Ma, and Song Guo. Wmpo: World model-based policy optimization for vision-language-action models. arXiv preprint arXiv:2511.09515, 2025. 3, 25 [51] Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv preprint arXiv:2506.09985, 2025. 3 [52]Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator, 2023. 3 [53]Xiao Fu, Xintao Wang, Xian Liu, Jianhong Bai, Runsen Xu, Pengfei Wan, Di Zhang, and Dahua Lin. Learning video generation for robotic manipulation with collaborative trajectory control, 2026. 3 [54] Fangqi Zhu, Hongtao Wu, Song Guo, Yuxiao Liu, Chilam Cheang, and Tao Kong. Irasim: A fine-grained world model for robot manipulation, 2025. 3 [55]Siyuan Huang, Liliang Chen, Pengfei Zhou, Shengcong Chen, Zhengkai Jiang, Yue Hu, Yue Liao, Peng Gao, Hongsheng Li, Maoqing Yao, et al. Enerverse: Envisioning embodied future space for robotics manipulation. arXiv preprint arXiv:2501.01895, 2025. [56] 1X World Model Team. 1x world model: Evaluating bits, not atoms. Technical report, 1X, 2025. [57] Wei-Cheng Tseng, Jinwei Gu, Qinsheng Zhang, Hanzi Mao, Ming-Yu Liu, Florian Shkurti, and Lin Yen-Chen. Scalable policy evaluation with video world models. arXiv preprint arXiv:2511.11520, 2025. 3 [58]Chenhao Li, Andreas Krause, and Marco Hutter. Robotic world model: A neural network simulator for robust policy optimization in robotics, 2025. 3 [59]Jiahan Zhang, Muqing Jiang, Nanru Dai, Taiming Lu, Arda Uzunoglu, Shunchi Zhang, Yana Wei, Jiahao Wang, Vishal M. Patel, Paul Pu Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan Yuille, Yilun Du, and Jieneng Chen. World-in-world: World models in a closed-loop world, 2025. 3 15 [60]Erick Rosete-Beas, Oier Mees, Gabriel Kalweit, Joschka Boedecker, and Wolfram Burgard. Latent plans for task-agnostic offline reinforcement learning, 2022. 3 [61]Rutav Shah, Shuijing Liu, Qi Wang, Zhenyu Jiang, Sateesh Kumar, Mingyo Seo, Roberto Mart ́ın-Mart ́ın, and Yuke Zhu. Mimicdroid: In-context learning for humanoid robot manipulation from human play videos, 2025. 3 [62]Simon Hangl, Emre Ugur, Sandor Szedmak, and Justus Piater. Robotic playing for hierarchical complex skill learning. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2799–2804. IEEE, 2016. 3 [63] Simon Hangl, Vedran Dunjko, Hans J. Briegel, and Justus Piater. Skill learning by autonomous robotic playing using active learning and creativity, 2017. [64]Irmak Guzey, Ben Evans, Soumith Chintala, and Lerrel Pinto. Dexterity from touch: Self-supervised pre-training of tactile representations with robotic play. arXiv preprint arXiv:2303.12076, 2023. 3 [65]Pulkit Agrawal, Ashvin V Nair, Pieter Abbeel, Jitendra Malik, and Sergey Levine. Learning to poke by poking: Experiential learning of intuitive physics. Advances in neural information processing systems, 29, 2016. 3 [66]Frederik Ebert, Chelsea Finn, Sudeep Dasari, Annie Xie, Alex Lee, and Sergey Levine. Visual foresight: Model- based deep reinforcement learning for vision-based robotic control. arXiv preprint arXiv:1812.00568, 2018. [67]Sudeep Dasari, Frederik Ebert, Stephen Tian, Suraj Nair, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Sergey Levine, and Chelsea Finn. Robonet: Large-scale multi-robot learning. arXiv preprint arXiv:1910.11215, 2019. [68]Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International journal of robotics research, 37(4-5):421–436, 2018. [69] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In 2016 IEEE international conference on robotics and automation (ICRA), pages 3406–3413. IEEE, 2016. 3 [70] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Devin, Alex X Lee, Maria Bauza, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving foundation agent for robotic manipulation. arXiv preprint arXiv:2306.11706, 1(8), 2023. 3 [71] Homer Rich Walke, Jonathan Heewon Yang, Albert Yu, Aviral Kumar, Jedrzej Orbik, Avi Singh, and Sergey Levine. Don’t start from scratch: Leveraging prior data to automate robotic reinforcement learning. In Conference on Robot Learning, pages 1652–1662. PMLR, 2023. 3 [72] Akshay L Chandra, Iman Nematollahi, Chen Huang, T. Welschehold, Wolfram Burgard, and Abhinav Valada. Diwa: Diffusion policy adaptation with world models. ArXiv, abs/2508.03645, 2025. 3, 25 [73]Zhiyuan Zhou, Pranav Atreya, Abraham Lee, Homer Walke, Oier Mees, and Sergey Levine. Autonomous improvement of instruction following skills via foundation models. arXiv preprint arXiv:2407.20635, 2024. 3 [74]Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX security symposium (USENIX Security 23), pages 5253–5270, 2023. 4 [75] OpenAI. Gpt-4 technical report, 2024. 5 [76] Physical Intelligence et al.π 0.5 : a vision-language-action model with open-world generalization, 2025. 5, 18 [77] Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, et al. Libero-plus: In-depth robustness analysis of vision-language-action models. arXiv preprint arXiv:2510.13626, 2025. 5 [78]Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail, 2020. 5 [79] Youguang Xing, Xu Luo, Junlin Xie, Lianli Gao, Hengtao Shen, and Jingkuan Song. Shortcut learning in generalist robot policies: The role of dataset diversity and fragmentation, 2025. 5 [80] Yoshua Bengio, J ́erˆome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48, 2009. 5 16 [81]Liang Wen, Yunke Cai, Fenrui Xiao, Xin He, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Tanglifu Tanglifu, Xiaowei Lv, et al. Light-r1: Curriculum sft, dpo and rl for long cot from scratch and beyond. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track), pages 318–327, 2025. 5 [82] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PmLR, 2021. 6 [83]Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7 [84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018. 8 [85]Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 8 [86] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, 44(10-11):1684–1704, 2025. 8 [87] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π 0 : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 8 [88]Lirui Wang, Kevin Zhao, Chaoqi Liu, and Xinlei Chen. Learning real-world action-video dynamics with heterogeneous masked autoregression, 2025. URL https://arxiv.org/abs/2502.04296. 10 [89]Zhengtong Xu, Qiang Qiu, and Yu She. Vilp: Imitation learning with latent video planning, 2025. URL https://arxiv.org/abs/2502.01784. 10 [90]Andrew Wagenmaker, Mitsuhiko Nakamoto, Yunchu Zhang, Seohong Park, Waleed Yagoub, Anusha Nagabandi, Abhishek Gupta, and Sergey Levine. Steering your diffusion policy with latent space reinforcement learning. arXiv preprint arXiv:2506.15799, 2025. 10, 22 [91] William Liang, Sam Wang, Hung-Ju Wang, Osbert Bastani, Yecheng Jason Ma, and Dinesh Jayaraman. Tether: Autonomous functional play with correspondence-driven trajectory warping, 2026. URLhttps://arxiv.org/ abs/2603.03278. 21 [92] Ansh Kumar Sharma, Yixiang Sun, Ninghao Lu, Yunzhe Zhang, Jiarao Liu, and Sherry Yang. World-gymnast: Training robots with reinforcement learning in a world model, 2026. URLhttps://arxiv.org/abs/2602.02454. 25 17 Appendix AImplementation Details Here we provide more implementation details ofPlayWorldthat expands upon Sec. 3, including the setup for play data collection (Sec. 3.2), and training details for the action-conditioned video model (Sec. 3.3). Data Collection Design. One ideal approach for data collection for training world models is to use an exploration policy that selects actions to maximally reduce the world model’s uncertainty (i.e., to intentionally seek out transitions that are most informative for the learned dynamics). While appealing in theory, we found this objective difficult to realize in a reliable, deployable system. In particular, using uncertainty estimates from a continually-updated video world model as a real-time reward signal can be unstable during robot execution, and can also create substantial operational overhead, since the model must be updated frequently as new data arrives, and the exploration signal can drift in ways that complicate monitoring and safety. Instead of directly optimizing “informativeness” through world-model uncertainty,PlayWorlduses a simple but robust proxy that maximizes the diversity of behaviors while staying close to the policy’s natural action manifold. Empirically, this produces broad coverage of interaction modes and state transitions without requiring a brittle closed-loop interdependence between exploration and model uncertainty estimation. Concretely, to make the system readily portable across tasks with minimal engineering, we instantiate our play collector withπ 0.5 -DROID [76]. Beyond its general instruction-following capability,π 0.5 -DROID reliably executes scene resets and handles a wide range of language prompts, which dramatically reduces human supervision, enabling persistent, scalable data collection. We provide further details analyzing the properties of the collected play data in Appendix B. Example Prompt You are a robot that is trying to randomly manipulate/arrange objects on the table in a square region marked by tape. Ignore anything that’s not on the gray table. First, observe if any object is outside of the square workspace area. If so, please output: “move the <> towards the center of the table." Otherwise, please come up with a short task for the robot to perform. Possible instructions might include: 1) “Put the <> on the <>." 2) “Remove the <> from the <>." 3) “Put the <> near the <>." 4) “Pick up the <> and put it onto the <>." Feel free to modify the prompt with different verbs and nouns (e.g. “put”→“move”, “block”→”object”). Use color to help the robot better identify the objects. Don’t include quotation marks in the task. Example In-Distribution Tasks Put the carrot in the white bowl. Remove the white plush toy from the bowl Pick up the orange rectangular block and put it on the gray table. Fold the towel from left to right. ... Example Out-of-Distribution Tasks Slide the carrot near the beige container. Move the orange rectangle onto the yellow cube. Flip the block up-side-down. Drag the blue towel across the table. ... 18 Figure 12 Example extracted centroids from the demonstration data that roughly corresponds to distinct nominal behavior episodes. Table 2 Hyperparameters used for video model training. HyperparameterValue Learning Rate5e-6 Warmup Steps100 Schedulercosine Gradient clipping1.0 Global batch size64 Observation Shape320 × 192 × 3 Num. Predicted Frames5 Num. History Frames7 Frame-level ConditionTrue Frequency5 Hz Denoising Steps50 Guidance2.0 Video Model Training. Our video model follows the setup of [9], which jointly predicts 3 camera views (left, right, and wrist views, each with resolution 192×320) given the current view and the Cartesian-space end-effector pose. We first pre- train the base SVD model on the full DROID dataset, and then fine-tune using the curriculum described in Sec. 3.3, with the hyperparameters in Table 2. The goal of this curriculum is to progressively shift training mass toward harder, rarely occurring transitions, preventing the model from overfitting to dominant free-space motion patterns. In practice, we compute a representation for each trajectory by concatenating CLIP embeddings from four sparsely-sampled frames over a 1-second window. For each task, we useK= 5 to identify nominal behavior modes via clustering, and setR= 5 ranks per task to construct the curriculum sampling buckets. We initialize the rank-sampling distribution as (0.5,0.3,0.1,0.05,0.05) and anneal it to (0.1,0.2,0.2,0.25,0.25) by uniformly updating the probabilities every 5000 steps. Fig. 12 shows the most representative sample (centroid) from each cluster that is used to construct the ranks, illustrating that the discovered modes align well with intuitive task-level sub-behaviors. While we used demonstration data as a proxy to define ranks with different motion attributes, we find that under this simplistic setup, the specific parameters used can have noticeable impact on model quality. Future works could investigate how to make the curriculum-like setup more generalizable for data at larger scale, where a more robust signal for motion quality could potentially be directly learned from data. BPlay Data Analysis Play data provides a scalable way to transform large volumes of a robot’s uncurated experience into a high-quality, simulatable environment, with minimal human involvement. In this section, we provide further analysis on properties of play data in comparison to human-collected alternatives to demonstrate that autonomously-collected data can exhibit meaningful behavioral diversity. Play Data Visualization. Fig. 13 provides qualitative snapshots from the play data corpus on a few different object sets, highlighting the natural variations that emerge when the robot explores without a scripted task objective (e.g., diverse object contacts, viewpoints, and interaction outcomes). Data Distribution at Different Scales. To complement these examples with a quantitative view, we include an extended t-SNE analysis (Fig. 14) across multiple dataset scales. Specifically, we embed the initial frame of each trajectory using CLIP features and visualize the resulting state-space coverage. Across matched data budgets, play data consistently spans a broader region of the embedding space than human-collected trajectories, while demonstration-based data concentrates in a comparatively narrow subset of states. As autonomously collected data scales beyond the practical limits of human collection, we expect it to cover the state space more densely and more broadly, yielding richer supervision and more informative transitions for 19 Figure 13 Example play data episodes selected from autonomous roll-outs for various tasks. Figure 14 Training sample distribution for demonstration data, human-collected play data, and autonomously-collected play data at different scales. We observe that play data consistently provides the widest distribution over the state space, while other human-collected data exhibits limited expansion in coverage even at larger scale. 20 Table 4 Per-category perceptual similarity across different training data mixtures. PSNR(↑) Training MixSuccessMissed GraspSlideSlipDeformationCollision Human Demo24.7824.3723.9823.6622.5824.67 Human Play23.5624.4123.9024.0122.9324.91 Robot Play24.9524.8325.0224.8423.2224.61 Robot Play (Curriculum) 25.0124.8225.1424.7823.5524.95 MSE(↓) Training MixSuccessMissed GraspSlideSlipDeformationCollision Human Demo279.80280.87341.36324.57428.55354.46 Human Play284.01276.54302.10332.65390.98304.12 Robot Play274.81 259.14263.17259.10366.55297.34 Robot Play (Curriculum)265.13264.19250.46243.51350.12295.53 training the world model. Concurrent Works. Concurrent with our work, [91] explores the use of autonomous play for policy learning by generating task-directed interactions guided by vision-language models (VLMs). Their results show that play-collected data can achieve competitive performance compared to human demonstrations, highlighting the potential of autonomous play as a scalable and sustainable source of training data for robot learning. CAdditional Trajectory Replay Results In this section, we provide more details on the trajectory replay experiments, including the composition of our evaluation benchmark, details on the scaling experiments, and full results on additional perceptual metrics. Evaluation Benchmark. Table 3 Composition of the trajectory replay benchmark. Source#Samples Human-collected97 π 0 85 π 0.5 122 Diffusion Policies229 Category#Samples Success98 Missed Grasp88 Slide76 Slip89 Deformation90 Collision92 Nominal task execution typically involves static contacts whose dynamics are comparatively consistent and easy to model. In contrast, dynamic object interactions such as collisions, deformation, or slip are substantially more diverse and harder to predict, yet they are often the decisive factors for counterfactual evaluation and task success. To target these failures, our benchmark is explicitly designed around predicting these interaction events. Since contact events are highly multi-modal, we curate the bench- mark from a large set of policy roll-outs and identify natural failure episodes that arise in practice. We then label each episode by failure mode, where the resulting distribution is summarized in Table 3. For every failure mode, our test set spans a wide range of trajec- tories. We argue that this coverage provides a more stringent and realistic assessment of action-conditioned video models, compared to typical nominal trajectory-replay evaluations which assess the model’s capability of generating realistic trajectories, rather than faithfully predicting effects of actions. For completeness, we provide additional results on the perceptual metrics in Table 4 (extending the results in Table 1). Scaling Experiments. In this section, we discuss the results in Fig. 10 in greater detail. To enable a fair and interpretable comparison across models trained with different dataset sizes, we convert the raw LPIPS metric 21 s raw into a normalized score s n using a fixed affine transform: s n =−(s raw − s min )/(s max − s min ).(7) Here,s min = 0.072 ands max = 0.10 are constants that define the LPIPS range used for normalization in our benchmark. This mapping preserves the ordering of models while (i) rescaling the narrow LPIPS interval to improve separation in scaling plots and (i) inverting the axis so that higher values correspond to better perceptual fidelity. As a result, scaling trends are easier to interpret, while all pairwise model comparisons remain unchanged. DAdditional Fine-Tuning Results Sim-to-real transfer is one of the most demanding end-to-end evaluations of a robotic simulator. Real robots expose compounded mismatch in contact, actuation, latency, and sensing, so policies trained in simulation often fail on hardware if they latch onto simulator-specific artifacts. In this section, we show that an accurate world model can allow RL fine-tuning with consistent improvement in real-world success rate, even on challenging fine-grained manipulation tasks. Stable RL Fine-tuning for Diffusion-Based Models. Although many modern robot policies are parameterized as diffusion or flow-matching models, RL fine-tuning these generators remains challenging: the policy is realized through an iterative sampling procedure, making it difficult to obtain stable, low-variance gradients with standard actor–critic objectives. Fine-tuning inside a learned world model introduces an additional source of brittleness, since effective exploration can hinge on the model capturing fine-grained action-dependent differences in predicted outcomes; small modeling errors can therefore distort the learning signal and lead to unstable updates. As a result, prior works primarily use world models as data generators, performing supervised fine-tuning (SFT) on filtered synthetic rollouts instead of directly optimizing returns. InPlayWorld, we show that when the world model is sufficiently accurate, direct RL fine-tuning can be surprisingly effective even for fine-grained manipulation, and can improve robustness beyond what is achievable from demonstrations alone. Fine-Tuning with DSRL. For our fine-tuning setup, we adopt the formulation in [90], which assumes a pretrained diffusion policy (e.g., trained with behavioral cloning) that generates an action sequence a conditioned on the current policy input s (observations and proprioceptive state). Sampling proceeds by drawing an initial latent-noise vector w∼N(0,I) and running a fixed reverse-diffusion denoiser to obtain the action: a = g θ (s,w),(8) whereg θ denotes the full reverse process (possibly multi-step) implemented by a frozen diffusion model with parameters θ. This induces the (marginal) policy distribution π θ (a| s) = Z δ(a− g θ (s,w)) N (w;0,I)dw,(9) i.e., actions are generated by pushing forward a simple noise prior through the denoiser. Following the “noise actor” formulation of Wagenmaker et al.[90], we replace the fixed priorN(0,I) with a learned, state-conditioned distribution over initial noise: w∼ π W φ (w| s), a = g θ (s,w),(10) where π W φ is a small policy (e.g., an MLP Gaussian) with parameters φ. This yields a steered action policy π φ,θ (a| s) = Z δ(a− g θ (s,w)) π W φ (w| s)dw,(11) which can be viewed as learning an expressive adapter in the diffusion model’s latent-noise space while leaving the diffusion parametersθunchanged. We optimizeπ W φ with an RL objective defined on environment rollouts 22 produced by executing a =g θ (s,w). Letτ= (s 0 ,w 0 ,r 0 ,s 1 ,...) denote the resulting trajectory; then the goal is max φ J (φ) = E τ∼(π W φ ,g θ )   X t≥0 γ t r t   .(12) Crucially, we treat w as the RL action and the (frozen) diffusion samplerg θ as part of the environment interface. In practice, this allows standard off-policy actor–critic updates using replay tuples (s t ,w t ,r t ,s t+1 ), without requiring backpropagation through the world dynamics or through the multi-step reverse-diffusion procedure. The specific hyperparameters used can be found in Table 5. Reward Model Design. To enable efficient policy improvement, we train a simple progress-based reward model from a small set of demonstrations. Given two time steps, the model predicts task progress, which we convert into a dense per-step signal by taking the difference in predicted progress between consecutive steps: r t = ˆp(s t+1 )− ˆp(s t ). Despite its simplicity, this reward signal leads to fast and stable learning in our setting: we often observe measurable behavioral improvements within the first 10–20 episodes. In contrast, sparse terminal 0/1 rewards require substantially more interaction to achieve comparable gains. We also explored vision-language-model (VLM) rewards, but found them less reliable in practice, often destabilizing fine-tuning with false predictions. Table 5 Hyperparameters used for DSRL training and diffusion policy. HyperparameterValue Obs dim8 Action dim8 Image shape[6, 192, 192] Cond steps (state)1 Cond steps (image)1 Horizon16 Max episode steps30 Actor LR1e-5 Batch size64 Discount0.99 Train freq15 UTD10 Tau0.001 Layer normTrue MLP backbone3 layers × 256 Target Entropy0.0 Initial rollout steps800 Action magnitude0.5 Denoising steps100 DDIM steps5 Prediction target ε Spatial embedding128 Experiment Setup. To show the power of fine-tuning withPlayWorld, we use diffusion policies trained from scratch with 5−10 demonstrations. We then collect 10−20 initial object configurations as environment initializations for the world model, and perform fine-tuning by randomly sampling among the available initializations. With each checkpoint, we perform 20 hardware trials to calculate the success rates shown in Fig. 9. Effectiveness of Fine-Tuning and World Model Hacking. Here, we report some interesting observations from the fine-tuning experiments. First, we find that stochastic predictions from the world model are beneficial 23 Figure 15 Task: remove the carrot from the bowl. Top: base policy behavior in the real world. Bottom:PlayWorld fine-tuned policy behavior in the world, where the robot learned to grasp a more stable position. Figure 16 Task: fold the towel from left to right. Top: base policy behavior in the real world. Bottom:PlayWorld fine-tuned policy behavior in the world, where the grasping motion is more robust and accurate. for learning robust policies, particularly under out-of-distribution (OOD) conditions. We hypothesize that stochastic video rollouts effectively act as a form of data augmentation: by exposing the policy to diverse, plausible futures during training, the resulting behavior generalizes to initializations that were not observed during either pretraining or fine-tuning. As shown in Fig. 16, when initialized from a novel configuration, the fine-tuned policy is able to gradually adapt its actions and eventually achieve a successful towel grasp. Second, we find that world-model hallucinations are not necessarily as detrimental to fine-tuning as one might expect. In several baseline fine-tuning runs, the learned policy partially collapsed into exploiting spurious model predictions—executing actions that are suboptimal (or nonsensical) in reality, yet yield high reward under the learned dynamics (i.e., “reward hacking” within the model). Despite this behavior in imagination, the resulting policies often still transferred well to the real robot. We hypothesize that fine-tuning can simultaneously (i) extract genuinely useful signal from the world model and reward, but may (i) overfit to hallucinated failure modes that are unlikely to occur under real dynamics; when deployed, these spurious gradients may have limited effect because the corresponding states are never realized. Overall, this suggests that meaningful performance gains are possible even with imperfect world models, provided that hallucinations do not dominate the policy updates that matter on real trajectories. Finally, we also observe the converse phenomenon: policies can overfit to the world model and degrade with prolonged fine-tuning. Because even small prediction errors (e.g., in depth, contacts, or control response) can shift the boundary between “success” and “failure” in the model, the set of trajectories that are scored as successful in imagination does not perfectly align with real-world outcomes. As a result, extended fine-tuning 24 can gradually bias the policy toward exploiting these mismatches, yielding apparent gains under the model while reducing real-world success rate. This effect is most pronounced when the policy repeatedly revisits a narrow subset of imagined high-reward states, effectively specializing to idiosyncrasies of the learned dynamics rather than improving the underlying task strategy. Practically, this motivates using conservative stopping criteria (e.g., early stopping on real rollouts) and/or regularization toward the pretrained policy to prevent overfitting to model errors. Related/Concurrent Works. Prior work on adapting video-based world models for RL-finetuning largely falls into two categories. The first trains task-specific world models, fine-tuned on narrow, single-task datasets to optimize performance for a predefined objective [50,72]. While effective within scope, these models often overfit to a limited interaction distribution and lack generality beyond the target task. The second direction explores generalist world model fine-tuning, where large pretrained video models are adapted using broad, internet-scale or multi-task data [92]. These approaches typically improve high-level capabilities such as language following, semantic consistency, or robustness to distractors, but do not explicitly optimize for contact-rich physical fidelity required for fine-grained visuo-motor control.PlayWorldinstead explores an intermediate space by collecting lab-scale interaction data that spans a broad distribution of contact events beyond any single task. We show that the resulting world model more faithfully captures fine-grained object motion and contact dynamics, enabling policy improvements that transfer reliably to the real world—yielding smoother execution, more accurate and robust end-effector control, and higher success rates on fine-grained manipulation tasks. EFinal Remarks We believe large-scale play data is a particularly promising and scalable supervision source for learning video world models, which can make learned simulators substantially more practical for policy development. In our lab setting, we were able to collect many hours of diverse, unlabeled interaction data with relatively low overhead, demonstrating its utility for various post-training applications. Looking forward, scaling data collection across multiple institutions and hardware setups could further expand coverage of environments, object sets, and embodiments, improving both the fidelity and the robustness of learned dynamics. Such a shared play-data ecosystem would enable us to build stronger, more generalizable world simulators that will be a valuable resource to the robot learning community. 25

Instant research discovery