Deploying generative World-Action Models for manipulation is severely bottlenecked by redundant pixel-level reconstruction, O(T) memory scaling, and sequential inference latency. We introduce the Causal Latent World Model (CLWM), which employs DINOv3 features as generative targets to disentangle interaction semantics from visual noise, yielding highly robust domain generalization. To overcome memory scaling, CLWM features a Dual-State Test-Time Training (TTT) Memory that guarantees a strict O(1) footprint for long-horizon tasks. To overcome deployment latency, we propose Speculative Asynchronous Inference (SAI) to mask partial diffusion denoising behind physical execution, cutting blocking latency by about 50%. To scale robust policies, we present EmbodiChain, an online framework that establishes the Efficiency Law by injecting an infinite flow of physics-grounded trajectories during training. Extensive experiments validate that CLWM achieves state-of-the-art performance in complex dual-arm simulation and unprecedented zero-shot sim-to-real transfer on physical robots, outperforming baselines explicitly finetuned on real-world data.
Figures & tables
Figure 1: Overview of the Causal Latent World Model (CLWM). CLWM employs a Mixture of Transformers (MoT) architecture that unifies a latent video model and an action model. To maintain historical context across interleaved latent frame and action tokens, a shared Test-Time Training (TTT) memory module dynamically updates its hidden states at flow time s=0 (working memory for action generation) or arriving new observations (long-term memory). With the latent video features obtained by a frozen DINOv3 model, CLWM performs autoregressive generation: the latent video model first predicts future features via flow matching, which subsequently conditions the action model to decode the corresponding action chunks.
Figure 2: Architecture of the TTT Memory Module. (a) Standard causal attention relies on a KV cache to maintain historical context. (b) Our architecture replaces the KV cache with a Test-Time Training (TTT) Layer. (c) The Dual-State TTT Memory Update Strategy. We maintain a Long-Term TTT Memory updated exclusively by real historical observations. For each generation step, a Working (Short-Term) TTT Memory is forked from the long-term weights to condition the latent video generation. It remains frozen during the ODE integration but is immediately updated at flow time s=0 using the predicted latent state.
Figure 3: The Speculative Asynchronous Inference Pipeline. (a) Conventional autoregressive pipeline incurs high blocking latency by strictly waiting for the action execution and the true sensor observation ot+1 / ft+1 to arrive before next-step generation. (b) SAI leverages predicted future semantics f^t+1 to proactively perform pre-denosing in the background. Upon observation concluding, new historical context are updated in TTT Memory modules, and only minimal fine-grained denoising is required to trigger the final predicted action a^t+1 . This effectively reduces the inference latency during deployment.
Figure 4: Schematic illustration of the Efficiency Law: loss as a function of the rate of data generation.
Figure 5: Articulated 3D objects generated by predicting a part-decomposed structure, synthesizing part geometry and appearance, and estimating articulation parameters for physics-based simulation ( Liu et al., 2026 ) .
Figure 6: Example of a generated scene layout for robot learning environments, illustrating the placement of interactive objects and background assets to ensure a collision-free, physically plausible layout.
Figure 7: Robot workspace visualization.
Simulation Task
π0.5
X-VLA
Motus
LingBot-VA
Ours
Adjust Bottle
99%
99%
93%
94%
100%
Beat Block Hammer
93%
88%
88%
98%
98%
Blocks Ranking RGB
85%
83%
97%
98%
98%
Blocks Ranking Size
26%
74%
63%
96%
97%
Click Alarmclock
89%
99%
100%
100%
100%
Click Bell
66%
100%
100%
100%
100%
Table 1: Success rates on the RoboTwin benchmark across various tasks. Our method (CLWM) demonstrates superior performance compared to previous baseline policies.
Configuration
ID Success (%)
OOD Success (%)
Baseline (Spatial Randomization Only)
64%
25%
+ Visual Augmentation
75%
42%
+ Physics-grounded Generation
81%
56%
+ Reachability-aware Sampling (Full)
95%
82%
Table 2: Ablation study on EmbodiChain domain expansion techniques. Average success rates (%) are reported under In-Distribution (ID) and Out-of-Distribution (OOD) conditions (evaluating over unseen objects, textures, lighting, and spatial layouts).
Training Configuration
Hanging Mug
Turn Switch
Stack Bowls
Static Baseline (1,500 demos)
62%
85%
88%
ODSsample 213
60%
84%
85%
ODSsample 50
92%
92%
96%
ODSsample 10
96%
98%
98%
Table 3: Ablation studies on Online Data Streaming (ODS) vs. static dataset training. All configurations use 5,000 training iterations and a batch size of 64. A lower replay bound in ODS indicates higher throughput of novel experiences.
Methods
Dual-Arm Water Pouring
Table Rearrangement
Items Hand-Over and Place
Pan Open and Place
π0
25%
20%
20%
5%
GR00T N1.5
35%
20%
15%
5%
Sim2Real-VLA
80%
80%
40%
35%
CLWM (Ours)
95%
90%
80%
65%
Table 4: Real-world deployment success rates across four challenging manipulation tasks, evaluated zero-shot using only simulation training data. Results are reported as success fraction ± standard error.
Embodied world models learn to predict future physical dynamics from visual observations and control signals, where physical knowledge is implicitly entangled within latent representations. We introduce CausalWM, a 16B embodied world model that performs explicit causal chain-of-thought reasoning before future video prediction. CausalWM organizes useful variables into a reasoning trajectory, allowing the model to progressively capture causal dependencies underlying physical evolution. To train CausalWM, we collect 31K hours embodied data and develop a three-stage paradigm consisting of large-scale video pre-training, causal CoT mid-training, and multi-objective RL post-training. Despite using only a limited set of supervised CoT variables, CausalWM exhibits emergent in-context learning capabilities, enabling contextual visual feature guidance and efficient few-step generation. CausalWM achieves state-of-the-art performance across language-conditioned, action-conditioned, single-view and multi-view benchmarks, including Top-1 performance on TriWorldBench leaderboard.
Ziming Xu, Shuang Liang, Ruobing Han +10
Aether AI · University of California, San Diego · Vanderbilt University
Dexterous manipulation demands long-term foresight and rapid reactive control. Vision-Language-Action (VLA) models, while proficient in high-level reasoning, often lack a fine-grained understanding of physical dynamics and spatial perception. Conversely, World-Action Models (WAMs) typically suffer from high inference latency due to iterative generation. These deficiencies result in a critical temporal misalignment where the model's intent fails to adapt to rapid physical contact changes. To overcome this fundamental bottleneck, we propose LiMA, an asynchronous dual-system generative framework that systematically decouples intent planning from reactive execution. LiMA organizes computation into a multi-scale hierarchy: a slow system handles sparse long-horizon spatiotemporal intent generation, while a fast system focuses on dense high-frequency motion refinement. To align sparse intent predictions with dense action trajectories, we introduce a Latent Schrödinger Bridge Coupling mechanism that formulates refinement as an entropy-regularized probabilistic transport process. LiMA reduces inference latency by 45.8% compared with Cosmos-Policy via asynchronous decoupling. Evaluated across six bimanual dexterous manipulation tasks spanning multiple horizons, LiMA achieves an overall success rate of 70.8% and an average subtask success rate of 78.9%, while maintaining performance in unseen scenarios. The project website is available at https://ccdcs.github.io/LiMA_repo/
Ning Chen, Yankai Fu, Junkai Zhao +5
State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · Beijing Academy of Artificial Intelligence
World action models (WAMs) improve robot control by modeling how observations evolve, but generating future observations at test time incurs substantial latency. Fast-WAM removes this process for efficiency; however, our matched implementations show lower generalization for Fast-WAM than for future-aware alternatives, especially with scarce robot demonstrations and in out-of-distribution scenarios. To bridge this gap, we introduce LAWA, a WAM architecture that uses compact latent actions as an operational representation of future intentions, enabling efficient test-time future imagination without generating future observations. Specifically, a discrete tokenizer enhanced by action-free pre-training produces manipulation-centric codebook targets. LAWA jointly denoises a continuous latent state anchored to these targets with executable action chunks while omitting the future-video branch at inference. On RoboCasa, LAWA achieves state-of-the-art average success rates of 65.6% and 80.8% in the few-shot and full data settings, improving over the matched Fast-WAM baseline by 9.6 and 4.5 points, respectively. It also preserves the performance level of the matched Joint-WAM variant while requiring 42.9% lower inference latency. LAWA also demonstrates competitive zero-shot robustness on LIBERO-Plus and superior performance on real-world tasks. These results show that future imagination need not be discarded: retaining it with compact latent actions yields an effective trade-off among performance, generalization, and latency. Code and models will be released.