Action-chunking policies predict future controls but execute only a prefix before replanning. How should a fixed action-token budget be organized to support execution? We propose LeapBot-JAC, which allocates sparse future-control references and dense execution targets within one shared action expert. Both chains are supervised by demonstrated controls and jointly denoised at aligned physical-time positions. Directed attention allows execution tokens to read reference tokens subject to the host's visibility constraints. JAC adds neither action tokens nor parameters and retains the host's world-modeling process; both chains participate in inference, while only the execution chain drives the robot. On LIBERO, JAC reaches 98.3% success with FastWAM and LeapBot-WA, up from 97.0% and 97.3%. On zero-shot LIBERO-Plus, their success rates improve from 53.5% to 70.4% and from 73.1% to 77.3%, respectively. RoboTwin and two real-world tasks provide additional deployment evaluations. Single-training-seed comparisons and FastWAM ablations support action-role organization as a practical design choice under a fixed action-token budget.
Figures & tables
Figure 1: Motivation and budget allocation. (a) A conceptual illustration of how future placement constraints may affect current grasp choices; it is not an experimental rollout or a validated feasible solution. (b) JAC reallocates 32 action tokens into sparse reference and dense execution targets on the same control-time axis. Only the execution prefix drives the robot.
Figure 2: Overview of LeapBot-JAC. A shared action expert jointly denoises a sparse reference chain R and a dense execution chain A within a fixed action-token budget. Physical-time encoding aligns the two chains, while a directed attention mask allows A to attend to R and blocks direct reverse access, subject to the host’s visibility constraints. Both chains are supervised by demonstrated controls, and only the execution prefix drives the robot. The illustrated K=10 is used on LIBERO/LIBERO-Plus; RoboTwin uses K=16 . The host’s world-modeling objectives and sampling schedule are preserved.
Figure 3: Time alignment and directed visibility. (a) At s=3 , reference indices span 0–45 and execution indices span 0–15. (b) The outlined 16×16 block prevents R queries from reading A keys. Sc,Sf denote current/future semantics. The full matrix illustrates an action-independent semantic context; non-action restrictions are inherited from the host. General visibility follows Eq. ( 8 ).
LIBERO
RoboTwin 2.0
Model
Spa.
Obj.
Goal
Long
Avg.
Clean
Random
Avg.
Literature reports
π0 ( Yuan et al., 2026 )
96.8
98.8
95.8
85.2
94.1
65.92
58.40
62.16
π0.5 ( Yuan et al., 2026 )
98.8
98.2
98.0
92.4
96.9
82.74
76.76
79.75
Motus ( Bi et al., 2025 )
96.8
99.8
96.6
97.6
97.7
88.66
87.02
87.84
LingBot-VA ( Li et al., 2026a )
98.5
99.6
97.2
98.5
98.5
92.93
91.55
92.24
Table 1: Standard manipulation. Success rates (%) on LIBERO and RoboTwin 2.0. Cited results and reproduced Base/JAC pairs are separated. Parentheses denote the improvement in average success rate (pp); bold indicates the better result within each reproduced pair. Our RoboTwin experiments execute 16 actions from each predicted action chunk; literature protocols are retained as reported. Motus LIBERO scores are reported by Yuan et al. (2026) ; its RoboTwin scores follow Bi et al. (2025) .
Model
Cam.
Robot
Lang.
Light
Bg.
Noise
Layout
Agg.
Literature reports
π0 ( Li et al., 2026b )
13.8
6.0
58.8
85.0
81.4
79.0
68.9
53.6
LAWA ( Li et al., 2026b )
69.2
66.7
62.8
96.2
64.5
85.5
78.6
74.4
ImageWAM (FLUX.2 4B) ( Zhang et al., 2026 )
80.8
50.3
91.4
98.1
85.5
93.8
80.5
83.1 †
Our reproductions
FastWAM
19.2
46.2
69.9
82.0
53.7
41.0
62.8
53.5
Table 2: Zero-shot LIBERO-Plus. Success rates (%) without Plus adaptation. Aggregate is full-test Total for our rows and trial-level aggregation for π0 /LAWA; † retains ImageWAM’s reported Avg.; equivalence to our episode-weighted Total is not assumed. Parentheses denote changes in aggregate success rate (pp); bold indicates gains.
Information flow
A reads R
R reads A
Spatial
Object
Goal
Long
Avg.
R↔A
√
√
98.0
99.4
97.0
95.4
97.5
R→A
√
×
98.0
99.6
98.2
97.2
98.3
No direct cross-chain
×
×
98.6
100.0
98.0
94.2
97.7
A→R
×
√
98.4
99.6
98.0
96.6
98.2
Table 3: G1: visibility on LIBERO at s=3 , λR=0.2 . Permissions refer to direct attention before host intersection. Padded targets are loss-masked but remain attention-visible.
Figure 4: LIBERO sensitivity on FastWAM. (a) Reference-supervision weight at s=2 , with the no-JAC baseline shown as a separate dashed reference. (b) Reference stride at λR=0.2 ; the observed range is 0.3 pp. Measured settings share a restricted vertical scale; lines aid readability. Results use one training seed, without error bars.
Backbone
Configuration
Total params
Denoising
Latency
Peak memory
(M)
steps
(ms)
(GB)
FastWAM
Base ( 32A )
11701.62
10
337.0
25.00
JAC ( 16R+16A )
11701.62
10
337.7
25.00
LeapBot-WA
Base ( 32A )
8428.11
10
438.6
26.13
JAC ( 16R+16A )
8428.11
10
435.1
26.13
Table 4: Computational cost within each host. Model-call latency is wall-clock p50, including host-side scheduling and CPU–GPU transfers within inference. Measurements use the training hardware and configuration. Memory is peak allocated GPU memory.
Figure 5: Real-world task sequences, ordered left to right. (a) Apple-to-green-basket placement precedes green-block-to-blue-basket placement. (b) Drawer closing with possible cabinet displacement during contact.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Evaluation
Host
Stride
Reference weight
Main performance
FastWAM; LeapBot-WA
3
0.2
G1 visibility
FastWAM
3
0.2
G2 weight
FastWAM
2
0.1, 0.2, 0.5, 1
G3 stride
FastWAM
2, 3, 4
0.2
Appendix
Table 5: Reported JAC configuration groups. All use FastWAM unless another host is listed. All results use one training seed.
Parameter
Setting
Host training hyperparameters
Original FastWAM / LeapBot-WA recipes
Training and evaluation hardware
Eight NVIDIA H100 GPUs
Training seeds
One common seed across local experiments
Action-token budget
Base: 32A ; JAC: 16R+16A
Default reference settings
s=3 , λR=0.2 , R→A
Weight sweep
s=2 ; λR∈{0.1,0.2,0.5,1}
Appendix
Table 6: Implementation and evaluation settings. Host-specific hyperparameters follow the cited original papers; execution settings below apply to our paired comparisons.
World Action Models (WAMs) have emerged as a powerful paradigm for embodied intelligence, yet the prevailing reliance on pixel-level video generation creates a fundamental bottleneck. Forcing models to reconstruct task-irrelevant visual details dissipates representational capacity and renders policies vulnerable to visual distractors. In this paper, we propose LeapBot-WA, which establishes a novel Predictive-Latent paradigm for WAMs by operationalizing the Joint-Embedding Predictive Architecture (JEPA) as a World-Anchor. Departing from the traditional reliance on visual synthesis, LeapBot-WA shifts the core of world modeling to Predictive Semantic Alignment, extracting abstract physical dynamics directly within a latent foundation space. To bridge the modality gap between non-Gaussian predictive features and diffusion priors, we introduce the Isotropic Semantic Autoencoder (ISAE), which reshapes the anchor's latent space into a diffusion-friendly manifold to prevent off-manifold drift. Furthermore, we design an Asymmetric Mixture-of-Transformers (MoT) architecture. During training, an Anchor Diffusion Transformer acts as a privileged dynamics expert to guide the Action Diffusion Transformer; at inference, this heavy dynamics branch is pruned, enabling zero-overhead execution. LeapBot-WA achieves state-of-the-art performance among predictive models on LIBERO and matches top-tier generative WAMs on RoboTwin 2.0 without requiring large-scale trajectory pre-training. It further demonstrates superior zero-shot robustness to unseen environments and successful real-world transfer, establishing a highly efficient and robust latent-centric paradigm for scalable robotic control. Code: https://github.com/LeapWM/leapbot-wa.
Pei Liu, Nan Zheng, Lang Zhang +8
1The Hong Kong University of Science and Technology (Guangzhou) · 2The Hong Kong University of Science and Technology · 3Leapmotor +1
Generalist robot policies aim to map multimodal observations and linguistic task instructions to actions across diverse tasks. However, existing methods typically represent the future as a fixed, short video-action chunk. This short-term future captures local scene evolution for action execution, but it does not explicitly describe the stage-level future that specifies how a task should progress from its current stage to the next. We therefore distinguish two complementary futures for robot manipulation: a short-term physical future to capture local scene evolution and a stage-level semantic future to represent task progress. We introduce StageWAM, which augments a Motus-based World Action Model (WAM) with Stage-JEPA, a goal-conditioned Joint-Embedding Predictive Architecture (JEPA) predictor. Given the current observation and task instruction, Stage-JEPA uses a frozen V-JEPA2 encoder to extract the current-state representation and predicts the latent target of the next inferred stage. Across 50 RoboTwin 2.0 tasks in clean and randomized environments, StageWAM achieves 90.25% overall success and reduces the mean number of execution steps in successful rollouts by 5.97% relative to the strongest baseline.
Xiao Liu, Yuguang Yang, Xi Wang +6
Institute for AI Industry Research (AIR), Tsinghua University · School of Electronic Information Engineering, Beihang University · AIR Wuxi Innovation Center, Tsinghua University +3
Robust robot control benefits from explicitly modeling state transitions, but video-generation world action models (WAMs) introduce substantial deployment cost. Existing latent WAMs avoid explicit future generation, but often compress predictive representations or separate predictive modeling from the representations used for action generation. We introduce JEPA-WAM, a latent WAM built in a pretrained V-JEPA space, which couples latent transition prediction with continuous action generation through a shared predictor. JEPA-WAM predicts a spatially structured joint current-future target that captures task-shared visual temporal structure between current and future observations, while preserving dense patch-level correspondence. Through the shared predictor, transition supervision directly shapes the backbone, from which dedicated representations are extracted for action prediction. The same design can also be instantiated in pretrained VLA policies while preserving their original perception and action pathways. On LIBERO-Plus, JEPA-WAM achieves 79.2%, the best result without large-scale robot-policy pretraining, while its pretrained π0.5 instantiation reaches 86.3%, achieving the best overall performance. Experiments on RoboTwin 2.0 and real-world bimanual manipulation further demonstrate strong generalization under visual and spatial shifts.
Yihan Lin, Jiawei He, Shifeng Bao +6
School of Information, Renmin University of China, Beijing, China · XYZ Embodied AI, Beijing, China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China +3