1X World Model Challenge

• 1X World Model Challenge is a benchmark for generative world models applied to real humanoid-robot interaction data, predicting future visual observations based on robot state and action history.
• It features two tracks—sampling for future image frame prediction and compression for discrete token forecasting—with winning scores of 23.0 dB PSNR and 6.6386 Top-500 CE.
• The challenge integrates robot-state conditioning with spatio-temporal modeling and ensemble inference to enhance predictive accuracy and simulate embodied control scenarios.

The 1X World Model Challenge is a benchmark for learning generative world models from real humanoid-robot interaction data, centered on predicting what a humanoid robot will observe in the future conditioned on its own state and action history. It is organized around two complementary tracks: sampling, focused on forecasting future image frames, and compression, focused on predicting future discrete latent codes. In the challenge’s winning technical report, Team Revontuli adapts Wan-2.2 TI2V-5B to video-state-conditioned future frame prediction and trains a Spatio-Temporal Transformer from scratch for discrete-token forecasting, achieving 23.0 dB PSNR in the sampling task and a Top-500 CE of 6.6386 in the compression task (Mereu et al., 8 Oct 2025).

1. Conceptual scope and formal setting

The challenge sits within a broader research program in which world models are treated as internal simulators of future states. In contemporary formulations, a world model may be written as a next-state simulator $s' \sim p(s' \mid s, a)$, or as a parametric approximation of an action-conditioned transition kernel such as $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$ in a partially observable MDP (Team et al., 26 Mar 2026, Shang et al., 18 May 2026). Within that perspective, the 1X benchmark specializes the problem to real-world humanoid interaction, with visual observation prediction conditioned on proprioceptive robot state.

A plausible implication is that the benchmark targets a particularly demanding regime of embodied prediction: the model must account simultaneously for egocentric scene evolution, humanoid self-motion, and the induced visual consequences of robot-state trajectories. Unlike simulator-only or purely video-quality-oriented evaluations, the challenge is explicitly grounded in real-world humanoid robot interaction data and frames world modeling as a predictive interface for embodied control (Mereu et al., 8 Oct 2025).

2. Benchmark tasks, data, and representations

The benchmark uses clips containing 77 RGB frames at $512\times512$ together with a state sequence $\mathbf{s}_{0:76} \in \mathbb{R}^{77\times 25}$, where each state is 25-dimensional (Mereu et al., 8 Oct 2025). The two tracks differ primarily in the representation of future prediction.

Track	Input	Output	Evaluation
Sampling	First 17 frames $\mathbf{x}_{0:16}$ and $\mathbf{s}_{0:76}$	The 77th frame, 2 seconds into the future	PSNR
Compression	3 past token grids and $\mathbf{s}_{0:63}$	3 future token grids of $32\times32$	Top-500 CE

In the sampling track, the task is defined as: input the first 17 frames $\mathbf{x}_{0:16}$ and the whole state sequence $\mathbf{s}_{0:76}$, then predict the last frame in the 77-frame clip (Mereu et al., 8 Oct 2025). In the compression track, video frames are first encoded using the Cosmos $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$0 tokenizer. A block of 17 RGB frames at $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$1 becomes 3 token grids of shape $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$2, and the dataset contains 306k samples, each with 6 consecutive token grids and robot states $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$3 (Mereu et al., 8 Oct 2025).

The compression task is formalized as predicting future token grids from past token grids and robot states: $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$4 with training by teacher forcing to minimize

$\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$5

These two tracks therefore separate pixel-space forecasting from discrete latent-space dynamics modeling (Mereu et al., 8 Oct 2025).

3. Sampling track: adapting a foundation video model

For the sampling track, Revontuli adapt Wan-2.2 TI2V-5B, described as a large flow-matching generative video model with a 30-layer DiT backbone, to video-state-conditioned future prediction (Mereu et al., 8 Oct 2025). The starting point is a Text-Image-to-Video model, but the adaptation changes the conditioning regime from a single fixed image to a short past video segment together with robot-state trajectories.

The video is temporally downsampled by a factor of 4, turning the 77-frame clip into an effective 21-frame sequence. With the Wan2.2-VAE, the temporal latent length becomes

$\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$6

and the model operates on latent tensors of shape $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$7 (Mereu et al., 8 Oct 2025). The conditioning frames are the past frames preserved in latent form, while future latent slices are sampled.

Robot-state conditioning is implemented with AdaLN-Zero in every DiT block. The state pipeline first downsamples the state sequence to match the video schedule, augments continuous state values with sinusoidal features, projects them with an MLP to hidden size

$\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$8

and then compresses them temporally with a 2-layer 1D convolutional network to produce features of shape $\hat{\mathcal{P}}_\phi(o_{t+1}\mid o_t, a_t)$9 (Mereu et al., 8 Oct 2025). These features are mapped to modulation vectors that are added to the standard timestep modulation, so each latent time slice is conditioned on the aligned robot state.

To adapt the 5B-parameter model efficiently, the team uses LoRA with rank 32 inside the DiT backbone: $512\times512$0 Only the low-rank update is trained while the base model remains frozen (Mereu et al., 8 Oct 2025). Training uses AdamW with constant learning rate $512\times512$1, 23k training steps, and effective batch size 1024 on 4 nodes × 8 × NVIDIA B200 GPUs (Mereu et al., 8 Oct 2025).

At inference time, the report emphasizes ensemble averaging. Multiple stochastic generations $512\times512$2 are averaged,

$512\times512$3

which improves PSNR by smoothing regions of predictive uncertainty such as moving arms, though it worsens some perceptual metrics through over-smoothing (Mereu et al., 8 Oct 2025). The report explicitly compares this strategy with Gaussian blur + histogram matching and finds ensemble averaging more effective for the challenge metric.

4. Compression track: spatio-temporal token dynamics

The compression-track system is a Spatio-Temporal Transformer trained from scratch, described as being inspired by Genie and ViT (Mereu et al., 8 Oct 2025). Its input consists of past token grids together with robot-state embeddings, and its output is a categorical distribution over future token grids.

The design alternates spatial attention and temporal attention. Spatial attention is applied independently within each $512\times512$4 token grid, while temporal attention is applied causally across time at each fixed spatial coordinate. This factorization reduces memory cost relative to full joint space-time attention while preserving autoregressive temporal structure (Mereu et al., 8 Oct 2025).

The model configuration is explicit: 24 ST-Transformer layers, 8 attention heads, embedding dimension 512, Pre-LayerNorm, QKNorm, and dropout 0.1 for attention, MLP, and residual pathways (Mereu et al., 8 Oct 2025). Robot states are incorporated as additive embeddings obtained by projecting $512\times512$5 with an MLP, processing them with a 1D convolution with kernel 3 and padding 1, and adding absolute positional embeddings before fusion with token embeddings (Mereu et al., 8 Oct 2025).

The probabilistic factorization is written as

$512\times512$6

with training under teacher forcing and inference by greedy autoregressive decoding: $512\times512$7 The report notes that scheduled sampling was explored but did not significantly improve the gap between teacher-forced and true autoregressive performance (Mereu et al., 8 Oct 2025).

Optimization uses fused AdamW with $512\times512$8, $512\times512$9, weight decay 0.05, 2000-step warmup to peak learning rate $\mathbf{s}_{0:76} \in \mathbb{R}^{77\times 25}$0, then linear decay over 80 epochs (Mereu et al., 8 Oct 2025). Training uses AMP in bfloat16, but inference is performed in float32, because bfloat16 degraded performance (Mereu et al., 8 Oct 2025). A plausible implication is that numerical precision remains an important practical variable in discrete-token world modeling, even when the high-level architecture is comparatively compact.

5. Results and optimization behavior

The challenge report states that Revontuli secured 1st place in both challenges (Mereu et al., 8 Oct 2025). On the official leaderboard, the sampling track result is 23.00 dB PSNR on test, compared with 21.56 dB for the next-best submission, while the compression track result is Top-500 CE 6.64 on test, compared with 7.50 for the next-best submission (Mereu et al., 8 Oct 2025).

Track	Revontuli	Next best	Rank
Sampling	23.00 dB PSNR	21.56 dB PSNR	1
Compression	6.64 Top-500 CE	7.50 Top-500 CE	1

The sampling report also provides detailed validation behavior under different inference settings. With 20 inference steps and 20 ensemble samples, the model reaches validation PSNR 24.88 dB, SSIM 0.762, LPIPS 0.201, and FID 90.71 (Mereu et al., 8 Oct 2025). With 100 inference steps and 20 ensemble samples, validation PSNR reaches 25.53 dB, and a post-deadline test result of 23.04 dB is reported (Mereu et al., 8 Oct 2025). The best leaderboard submission trained on train+val reaches validation PSNR 26.62 dB, SSIM 0.836, LPIPS 0.082, and FID 31.70 (Mereu et al., 8 Oct 2025).

A consistent empirical pattern is that more aggressive ensembling improves PSNR while degrading LPIPS and FID because the resulting predictions become smoother (Mereu et al., 8 Oct 2025). This suggests a metric-specific optimization tension: if the objective is strictly pixelwise fidelity on a held-out future frame, model averaging is beneficial; if the objective were perceptual realism or downstream planning utility, the optimum could differ. The report explicitly identifies this tension as a limitation.

6. Position in the broader world-model benchmark ecosystem

The 1X World Model Challenge belongs to a rapidly expanding benchmark landscape in which world models are increasingly evaluated not only as predictors, but as simulators, planners, and evaluators.

Benchmark	Primary emphasis	Notable feature
WorldArena 2.0 (Shang et al., 18 May 2026)	Embodied world models across modality, functionality, and platform	Visuotactile prediction, world models as RL environments, sim-to-real evaluation
WR-Arena (Team et al., 26 Mar 2026)	Action Simulation Fidelity, Long-horizon Forecast, Simulative Reasoning and Planning	Evaluates world models as “thought experiment engines”
WorldMark (Xu et al., 23 Apr 2026)	Interactive Image-to-Video world models	Same scenes, same action sequences, unified control interface
WorldOlympiad (Zhao et al., 9 Jun 2026)	Physical faithfulness, geometric consistency, interaction fidelity	Gaming, robotics, and real-world long videos
WorldEval (Li et al., 25 May 2025)	World model as real-world robot policies evaluator	Policy2Vec turns a video generation model into a world simulator that follows latent action
WorldPrediction (Chen et al., 4 Jun 2025)	High-level world modeling and procedural planning	Action equivalents and counterfactual distractors
WorldModelBench (Li et al., 28 Feb 2025)	Video generation models as world models	Instruction following, physics adherence, commonsense

Within this ecosystem, the 1X benchmark is distinctive for at least three reasons. First, it is an open-source benchmark of real-world humanoid interaction with a direct emphasis on real humanoid-robot interaction data rather than simulator-only evaluation (Mereu et al., 8 Oct 2025). Second, it explicitly separates pixel-space forecasting and discrete latent forecasting, making representation choice itself part of the benchmark. Third, it places humanoid proprioceptive conditioning at the center of the task through a 25-dimensional state trajectory (Mereu et al., 8 Oct 2025).

This suggests that the challenge occupies a specific niche between interactive video-world benchmarks and embodied robot benchmarks. WorldArena 2.0 expands embodied evaluation along modality, functionality, and platform; WR-Arena emphasizes next world simulation and planning utility; WorldEval uses world models to rank real robot policies; WorldMark standardizes action-conditioned interactive video testing; WorldOlympiad decomposes evaluation into physical, geometric, and interaction tracks; WorldPrediction tests high-level action-state reasoning; and WorldModelBench formalizes physics and instruction-following judgments for generated videos (Shang et al., 18 May 2026, Team et al., 26 Mar 2026, Li et al., 25 May 2025, Xu et al., 23 Apr 2026, Zhao et al., 9 Jun 2026, Chen et al., 4 Jun 2025, Li et al., 28 Feb 2025).

A plausible implication is that the 1X challenge can be read as a humanoid-specific complement to these broader efforts: it emphasizes real embodied data, robot-state conditioning, and a dual representation view of world modeling, while leaving open the longer-term questions highlighted elsewhere in the literature—long-horizon rollouts, policy optimization inside learned worlds, simulative reasoning, physical faithfulness, and cross-platform transfer. The report itself identifies several open issues, including metric alignment versus perceptual realism, short-horizon focus, deployment cost of large pixel-space models, and generalization beyond the training distribution (Mereu et al., 8 Oct 2025).