Organizations: State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · The Chinese University of Hong Kong · Simplexity Robotics
World Action Models (WAMs) have recently emerged as a promising paradigm for robotic manipulation, endowing policies with strong capabilities for modeling temporal and physical dynamics. Existing WAMs, encompassing both video-first action prediction and joint video-action modeling, directly transfer global physical representations from world models to action generation. However, global scene representations may underemphasize local robot-object interactions, creating a mismatch between representations optimized for world evolution and those required for manipulation. To bridge this gap, we propose LaST0.5, a novel WAM that introduces an interaction-centric physical representation between world modeling and action generation. Specifically, LaST0.5 learns a compact interaction-centric latent space under the supervision of visual interaction cues between the robot and the environment. This reasoning space is instantiated as a latent expert that bridges the world expert and the action expert within a Mixture-of-Transformers architecture. Furthermore, LaST0.5 performs adaptive world-to-action transfer by jointly predicting future representations and progress values, using progress consistency to adjust action execution and trigger future resampling. Finally, we pretrain LaST0.5 on a large-scale robotic data corpus comprising over 14k hours of data, establishing interaction-centric physical reasoning as a scalable and effective paradigm. LaST0.5 achieves 99.5% on LIBERO, surpasses the previous state of the art on ten evaluated RLBench tasks, and demonstrates robust long-horizon execution and precise dexterous manipulation in real-world experiments.
Figures & tables
Figure 1: Overview. Existing WAMs directly transfer representations of global scene dynamics to action prediction, potentially overlooking the localized robot–environment interactions required for manipulation. LaST 0.5 bridges this gap through Interaction-Centric Latent Reasoning (ICLR), which learns a compact latent space supervised by visual motion and object-affordance cues, establishing a hierarchical reasoning paradigm that progressively transforms global world representations into interaction-centric abstractions and ultimately into robot actions. LaST 0.5 achieves state-of-the-art performance on the LIBERO and RLBench simulation benchmarks and demonstrates robust long-horizon and precise manipulation in real-world tasks.
Figure 2: The LaST 0.5 Framework. (a) LaST 0.5 unifies a World Expert, Latent Expert, and Action Expert within a Mixture-of-Transformers (MoT) architecture. To bridge global scene dynamics and precise local control, the Latent Expert distills future frames into compact, interaction-centric representations. (b) The Adaptive World-to-Action Transfer (AWAT) mechanism predicts progress values to assess task-progress consistency in imagined futures. Action execution is dynamically truncated when progress value drops, adapting the execution horizon to the predicted progress sequence. (c) Attention mask of LaST 0.5 .
Method
Close Box
Close Laptop
Toilet Seat Down
Sweep to Dustpan
Close Fridge
Phone on Base
Umbrella Out
Frame Off Hanger
Wine at Rack
Water Plants
Mean S.R. ↑ (mean ± std.)
OpenVLA-OFT ( Kim et al. 2025 )
100.0
65.0
60.0
30.0
80.0
30.0
30.0
20.0
20.0
15.0
45.0 ± 3.0
π0.5 ( Physical Intelligence et al. 2025 )
90.0
95.0
85.0
75.0
100.0
5.0
10.0
80.0
75.0
35.0
65.0 ± 4.0
HybridVLA ( Liu et al. 2025 )
85.0
95.0
100.0
90.0
100.0
50.0
50.0
70.0
50.0
50.0
74.0 ± 4.0
Cosmos Policy ( Kim et al. 2026 )
95.0
95.0
100.0
55.0
75.0
80.0
85.0
75.0
55.0
55.0
77.0 ± 2.0
Fast-WAM ( Yuan et al. 2026 )
95.0
95.0
95.0
75.0
100.0
25.0
75.0
60.0
0.0
10.0
63.0 ± 3.0
LaST 0 ( Liu et al. 2026 )
95.0
95.0
100.0
80.0
85.0
75.0
75.0
70.0
85.0
60.0
82.0 ± 3.0
Table 1: Comparison of LaST 0.5 and baselines on RLBench, reporting the mean success rate (%) across three independent runs.
Method
Spatial
Object
Goal
Long
Mean S.R. ↑
OpenVLA-OFT ( Kim et al. 2025 )
97.6
98.4
97.9
94.5
97.1
π0.5 ( Physical Intelligence et al. 2025 )
98.8
98.2
98.0
92.4
96.9
InternVLA ( Cai et al. 2026 )
98.6
99.8
98.6
98.4
98.9
Cosmos Policy ( Kim et al. 2026 )
98.1
100.0
98.2
97.6
98.5
Motus ( Bi et al. 2026 )
96.8
99.8
96.6
97.6
97.7
Fast-WAM ( Yuan et al. 2026 )
98.2
100.0
97.0
95.2
97.6
Table 2: Comparison on LIBERO, reporting the best success rate (%) obtained during training.
Figure 3: Ablation studies on (a) individual experts and (b) interaction-centric latent target construction, both using LaST 0.5 w/o AWAT as the “Base” model. (c) Temporal granularity for the value module, evaluated on the full LaST 0.5 .
Figure 4: attention maps of LaST 0.5 and its ablated variants.
Method
Bag Pulling
Object Organization
Screw Fastening
Parcel-Box Packing
Avg. ↑
S1
S2
S1
S2
S3
π0.5
75.0
85.0
80.0
30.0
95.0
75.0
20.0
52.5
Cosmos Policy
60.0
85.0
60.0
20.0
60.0
35.0
10.0
43.8
LaST 0
60.0
75.0
75.0
25.0
65.0
50.0
15.0
43.8
LaST 0.5 (Ours)
80.0
95.0
90.0
50.0
100.0
85.0
30.0
63.8
Table 3: Real-world results, reporting success rates (%) over 20 rollouts per task. Stage-wise breakdowns are provided for Screw Fastening (S1: Tool-use, S2: Insertion) and Parcel-Box Packing (S1: Place, S2: Fold, S3: Seal).
Figure 5: Example visualization of LaST 0.5 performing real-world tasks, with sequences shown from left to right.
Method
Scenario
Bag Pulling
Object Organization
Avg. ↑
π0.5
Position
15.0
15.0
15.0
Object
60.0
50.0
55.0
Background
60.0
75.0
67.5
Cosmos Policy
Position
0.0
35.0
17.5
Object
10.0
60.0
35.0
Background
60.0
65.0
62.5
Table 4: Generalization results on two real-world manipulation tasks. We report task-level success rates (%) under position, object, and background shifts.
Figure 6: Out-of-distribution scenario setting.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Full Attention Mask of the LaST 0.5 Architecture. The matrix visualizes the global attention mechanism across the three experts, where rows denote queries ( Q ) and columns denote keys ( K ). Colored cells indicate visible tokens (attention enabled), while white cells denote masked tokens (attention disabled).
Dataset
Modality
Episodes
Hours
Genrobot
UMI
656,993
9,088.0
AgiBot
Dual-arm
996,381
2,479.0
InternData-A1
Sim + Real
344,607
1,429.2
DROID
Single-arm
94,030
766.2
EgoDex
Human ego
314,839
727.4
RoboCOIN
Dual-arm
42,828
313.2
Appendix
Table 5: Composition of the LaST 0.5 pre-training corpus, sampled at 10 Hz. Sources: AgiBot ( Bu et al. 2025 ) , InternData-A1 ( Tian et al. 2025 ) , DROID ( Khazatsky et al. 2024 ) , EgoDex ( Hoque et al. 2025 ) , RoboCOIN ( Wu et al. 2025b ) , RoboMIND ( Wu et al. 2024 ) .
Figure 8: Real-World Robot Experiment Setup.
Latent tokens
Sampling
Success rate
3 (base)
t+4,8,12
99.6
6
t+2,4,…,12
99.4
12
t+1,2,…,12
99.3
Appendix
Table 6: Latent CoT capacity. We ablate the number of latent tokens by varying the temporal sampling density.
Latent target source
Success rate
SigLIP (base)
99.6
Cosmos, pre-DiT
99.1
Cosmos, post-DiT
99.4
Cosmos (action-conditioned), post-DiT
99.2
Appendix
Table 7: Latent target source. Performance comparison of different representations for latent targets.
Generation
Success rate
Parallel (base)
99.6
Autoregressive
99.4
Teacher forcing
97.5
Appendix
Table 8: Latent generation mechanism.
Figure 9: Action-to-vision attention maps.
Figure 10: Value progress visualization. The value curve acts as a real-time indicator of task status. It illustrates predicted task progress, rising during successful execution and dropping in the shown failure cases involving misaligned world knowledge.
Figure 11: Additional visualization of LaST 0.5 performing real-world tasks.
Figure 12: Illustrative failure sequences from four real-world tasks. Red bounding boxes indicate the exact instant and location where each execution error occurs.
World-action modeling has emerged as a promising paradigm for robotic control, as it empowers models to go beyond reacting to observations and anticipate how a scene will evolve. However, existing WAMs often incur substantial computational overhead. Pixel-space methods often allocate substantial capacity to visual details that may not be directly relevant to control, while some latent-space methods require multi-stage training to construct the reasoning space. The resulting training cost can make such methods difficult to train under modest computational budgets. In this work, we propose LiLa-WAM, a lightweight world-action model that reasons about the future in a compact latent space and can be trained end-to-end on a single 24GB GPU. Its core design is a compact latent reasoning space jointly shaped by future-state prediction and action generation, which keeps the model lightweight while remaining well aligned with control. For task specification, we further propose the Visual Transition Token(VTT), a language-free task representation that encodes each task as a direction in visual feature space. Experiments on RoboTwin~2.0, LIBERO, and real-robot tasks demonstrate LiLa-WAM's effectiveness, achieving 90.48% success across 50 RoboTwin tasks with single-GPU training.
Fan Yang, Yuting Su, Xiaobo Wang +7
1Tianjin University · 2Shenzhen University of Advanced Technology · 5Sangfor Technologies Inc. +2
World-Action Models (WAMs) build robot control on video-generation backbones, which jointly predict dense future visual trajectories and robot actions. We argue that video generation is an unnecessary intermediate objective for world-action modeling. For robotic manipulation, the goal of a world model is not to reproduce how the world looks at every intermediate moment, but to predict the state that the world will reach after an action is executed. The intermediate frames only describe the visual transition between physical states, which consumes substantial model capacity and computation, but do not directly specify the physical outcome that the robot action is intended to produce. In this paper, we propose DELE-w0.5, which infers robot actions from predicted future states without relying on video generation. Concretely, DELE-w0.5 infers the action sequence from its corresponding compact future latent state. The future latent state captures the action-relevant physical outcome of robot interaction and serves as an explicit bridge between world modeling and action generation. The core design principle of DELE-w0.5 is to model how the physical world changes under robot actions, rather than how its visual appearance evolves frame by frame. This formulation removes the high-dimensional visual redundancy introduced by dense video representations, and it therefore enables cheaper training and low-latency inference. Across 640 real-robot trials on four long-horizon manipulation tasks, our DELE-w0.5 achieves the best performance among all compared policies, attaining 62.5% overall full-task success and 81.3% macro ordered-stage progress. It outperforms the strongest baseline by 32.5 percentage points in full-task success and 20.1 percentage points in macro progress.
This work presents RepWAM, a representation-centric world action model (WAM) built on representation visual-action tokenizers. Existing WAMs typically inherit reconstruction-oriented video tokenizers from pretrained video generation models. Although these tokenizers preserve visual fidelity, pixel reconstruction alone provides limited guidance for learning instruction-following dynamics that connect future prediction with robot control. To address this, we explore a semantic visual-action latent space for representation-centric world action modeling. Specifically, we train a representation visual-action tokenizer that maps visual inputs into aligned visual and latent action tokens. We then pretrain our WAM to jointly model future visual states and the latent actions that connect them under language instructions, followed by adaptation to real robot trajectories for closed-loop manipulation. Experiments on real-world manipulation tasks and simulation benchmarks show that RepWAM delivers strong performance across diverse manipulation settings, while ablations highlight the value of semantic visual-action tokenization over reconstruction-oriented alternatives. These results establish representation visual-action tokenization as a promising foundation for world action models and a step toward generalist robot policies. Code and weights will be available at https://github.com/wdrink/RepWAM.
Junke Wang, Qihang Zhang, Shuai Yang +5
Institute of Trustworthy Embodied AI, Fudan University · 2Robbyant, Ant Group · 3Hongkong University of Science and Technology