1X World Model (1XWM) — action-controllable world model for humanoid policy evaluation

TL;DR

1X trains the 1X World Model (1XWM) as an action-controllable video predictor for its NEO humanoid: given a few initial frames plus a sequence of low-level robot action trajectories, it predicts future video frames and a scalar state-value head that estimates task success. The model is repurposed as a policy evaluator — instead of running each candidate robot policy in physical homes (intractable, slow, and unsafe), 1X rolls out policies inside 1XWM, scores them via the predicted success rate, and uses the rankings to make checkpoint and architecture decisions with same-day feedback. Reported behaviours include scaling of success-prediction accuracy with task-specific data, positive cross-task transfer (shelf+arcade improves shelf alone), and qualitative correlation between predicted and real success rates across policy ablations (proprioception input, ViT-L vs alternative image encoder). The post is a product/research blog from 1X (not the underlying technical report) and reports no aggregate benchmark numbers beyond a stated 70% accuracy → 90% correct-policy-pick rate at a 15% true gap.

Key claims

1XWM is trained on sequences of video frames, robot observations, and input action trajectories, encoding inputs to latent representations and predicting the latents of future frames plus a scalar state value of the final frame for success/completion [§ “training the world model on sequences…”].
Unlike text-to-video models, 1XWM is action-controllable: it is steered by exact robot trajectories rather than language directives, demonstrated by generating distinct futures from a shared initial state under four different action sequences (grab mug, wipe countertop, step back, play imaginary guitar) [§ Generation grids; § “Most video generation models are text-to-video…”].
Success-prediction accuracy improves with scale on both number of tasks and diversity of robot behaviours, demonstrated on Air Fryer, Arcade, and Shelf tasks [§ “Does 1XWM improve as data is scaled up?”].
Task-specific data fixes object-interaction failures: with smaller training corpora, 1XWM hallucinates the air-fryer tray and body as a single rigid unit; adding air-fryer interaction data lets it model tray-from-base separation and tray-confinement-within-base dynamics [§ “clear improvement in generation quality…”].
Positive cross-task transfer: training jointly on Shelf and Arcade data improves accuracy on Shelf compared to Shelf-only training [§ “training on both shelf and arcade…”].
1XWM is used as a policy evaluation engine: for each set of model-checkpoint weights, predicted success rates and futures are claimed to be distributionally aligned with real-world outcomes, enabling architecture and checkpoint decisions with an instant feedback loop [§ “An aligned world model can solve the evaluation problem…”].
Counterfactual evaluation on production states is explicitly listed as a capability — datasets of initial states where an autonomous policy previously failed are re-rolled out under candidate replacements [§ “curating datasets of production-setting states…”].
Ablation: 1XWM-predicted success rate correlates with real-world success when toggling proprioception (robot joint states) as input to the policy across checkpoints; the most- and least-promising 1XWM-ranked checkpoints reproduce that ordering on hardware [§ proprioception ablation plot].
Ablation: 1XWM correctly picks the better image-encoder option — the predicted-better ViT-L policy outperforms the alternative on real-world evaluation [§ “comparing two different image encoders…”].
Stated quantitative bound: given a true real-world success-rate gap of 15% between two policies, a world model with 70% accuracy can pick the better policy with 90% success [§ “Given a true real-world success rate gap of 15%…”].
Named limitation: 1XWM struggles to model interactions with held-out objects not seen in training (e.g. an unseen plant), making generalisation to novel objects a current bottleneck [§ “The 1XWM currently struggles…”].

Method

1XWM ingests sequences of robot video frames, robot observations (e.g. proprioception), and input action trajectories. Inputs are encoded into latent representations, and a network predicts the latent encodings of future frames; in parallel a head predicts the scalar state value of the final frame, used as a task success/completion estimator. The blog characterises this as predicting in latent space rather than directly in pixels, but does not disclose the tokenizer/VAE, the prediction backbone (autoregressive vs diffusion), training compute, parameter count, or context length. The action-controllability claim is operationalised by conditioning on a few real initial frames plus a sequence of explicit future robot poses/trajectories — the same initial frames are reused across multiple action sequences to demonstrate that the rollout obeys the action input rather than the visual prior alone (mug grab, countertop wipe, step backwards, imaginary guitar). Task-success training data appears to combine successful rollouts and failed attempts, so the value head learns a calibrated estimate rather than a video-quality proxy.

For policy evaluation, 1XWM is plugged in as the environment in a closed loop: a candidate policy checkpoint emits action trajectories, 1XWM rolls those out into predicted observation sequences and a final-state value, and predicted success rates are aggregated over a dataset of initial states. Two evaluation protocols are described: (i) production-setting datasets of initial states for in-distribution ranking, and (ii) counterfactual datasets seeded from states where the previously deployed autonomous policy failed, used to probe whether candidate replacements would have succeeded. Reported ablations use this pipeline to choose between policy variants (proprioception on/off, ViT-L image encoder vs alternative) before real-world deployment. No numerical correlation coefficient, calibration plot detail, or held-out test setup is given on the blog; the technical report referenced inline is not fetched here.

Results

Scaling: success-prediction accuracy improves across Air Fryer, Arcade, and Shelf tasks as the number of tasks and behaviour diversity grow [§ scaling plots]. No headline numbers reported on the blog.
Cross-task transfer: jointly training on Shelf + Arcade strictly improves Shelf-alone performance [§ transfer paragraph].
Object-physics specialisation: adding air-fryer data fixes the tray-as-rigid-body hallucination and induces tray-confinement-within-base dynamics qualitatively [§ air-fryer case study].
Policy ranking: predicted ranking over policy checkpoints correlates with real-world task scores under at least two ablations — proprioception input toggle and image-encoder choice (ViT-L vs alternative). The blog shows correlation plots but does not state an R² or absolute accuracy figure on the page.
Theoretical operating-point claim: at 70% WM accuracy and a 15% real-world success gap, the WM picks the correct better policy 90% of the time; the blog presents this as motivation, not as a measurement.
Failure mode: held-out objects (plant) produce visibly wrong interaction dynamics; this is acknowledged as a generalisation limit.

Why it’s interesting

This is the cleanest publicly described instance the wiki has of closed-loop policy evaluation in a humanoid-manipulation setting, and it is structurally identical to Evaluating Gemini Robotics Policies in a Veo World Simulator from Google DeepMind — both train an action-conditioned video predictor and use it to rank VLA policy checkpoints before hardware evaluation, with red-teaming / counterfactual variants as a use case. The 1X version predates the Veo Robotics writeup by more than a year (Oct 2024 blog vs Dec 2025 arXiv), so the wiki now has two independent labs converging on the same recipe — action-conditional WFM as the environment for policy evaluation, distinct from VJEPA-2 used as a reward for video generation in Inference-time Physics Alignment of Video Generative Models with Latent World Models.

It also tightens the empirical case for the World Foundation Models cluster’s “WFM as backbone” framing: 1XWM is not pitched as a generative product but as a tool that lets a robotics team make architecture decisions (proprioception, image encoder) before paying the cost of real-world rollouts. Compared to The Waymo World Model: A New Frontier For Autonomous Driving Simulation (closed flagship → autonomous-driving simulator) and Evaluating Gemini Robotics Policies in a Veo World Simulator (Veo → manipulation simulator), 1XWM is the rare first-party (not specialised from an external flagship) action-conditional WFM with disclosed policy-evaluation results, and from a startup rather than a major lab — a useful datapoint for the open question of whether the recipe needs a frontier video prior to work.