Fast-WAM
Removes future prediction at inference, reducing latency but weakening generalization in the matched comparisons.
current view→action
Efficient world action models
Latent Action as Intention Enables EfficientFuture Imagination for World Action Models
LAWA keeps the benefits of test-time future imagination while replacing expensivefuture-observation generation with compact, executable latent intentions.
1CollegeAI, THU2AIR, THU3CASIA4TARS Robotics5FDU6Southeast University7SJTU8NUS9BUAA
65.6%RoboCasa few-shot success
80.8%RoboCasa full-data success
74.4%LIBERO-Plus zero-shot success
42.9%Lower latency than Joint-WAM
Abstract
Imagine the future without rendering it.
Future imagination need not be discarded for efficiency: LAWA moves it from observation space into a compact latent-action space.
World action models improve robot control by modeling how observations evolve, but generating future observations at test time incurs substantial latency. Matched implementations show lower generalization for Fast-WAM than future-aware alternatives, especially with scarce demonstrations and out-of-distribution scenarios.
LAWA uses compact latent actions as an operational representation of future intentions, enabling efficient test-time future imagination without generating future observations. A discrete tokenizer enhanced by action-free pre-training produces manipulation-centric codebook targets; LAWA jointly denoises a continuous latent state anchored to those targets with executable action chunks.
Across RoboCasa, LIBERO-Plus, and real-world tasks, LAWA improves performance over Fast-WAM while preserving the performance level of Joint-WAM at substantially lower inference latency.
The design space
Three ways to connect perception and action.
Joint-WAM
Explicitly imagines future observations and feeds them to the action expert, preserving a direct, observation-space representation of task progress. This future awareness comes at the cost of iterative visual generation during inference.
current view→future pixels→action
LAWA
Predicts compact, temporally structured latent actions as intentions, then jointly denoises those intentions with action chunks.
current view→latent intention→action
Method
Joint attention turns latent transitionsinto executable control.
Training connects video, latent-action, and action experts. Inference removes the future-video branch and keeps only the compact intention pathway plus action prediction.
Tokenize transitions
A discrete tokenizer compresses current-to-future visual changes into a sequence of latent actions.
Predict intentions
LAWA denoises a continuous latent state anchored to manipulation-centric codebook targets.
Execute actions
The action expert attends to the latent intention and jointly denoises executable action chunks.
Action-free pre-training
A manipulation-centric tokenizer.
Video reconstruction can favor static appearance over small interaction dynamics. LAWA adds automatically generated mask targets to bias the tokenizer toward hands, manipulators, and interaction regions.
- Pre-trains on action-free robot and egocentric videos
- Quantizes temporally structured visual transitions
- Uses observation and auxiliary mask prediction
Results
A favorable performance–latency trade-off.
The comparisons answer complementary questions: Fast-WAM is the primary performance baseline,while Joint-WAM is the efficiency reference for explicit future generation.
Overall comparison on RoboCasa
Average success rate across 24 tabletop tasks.
| Method | Paradigm | Few-shot | Full data |
|---|---|---|---|
| GR00T N1.6 | VLA | — | 47.6 |
| StarVLA | VLA | — | 48.8 |
| StarVLA-α | VLA | — | 53.8 |
| TwinBrainVLA | VLA | — | 54.6 |
| ABot-M0 | VLA | — | 58.3 |
| RLDX-1 | VLA | — | 58.7 |
| JoyAI-RA | VLA | — | 63.2 |
| DIAL | VLA | 58.3 | 70.2 |
| Being-H0.7 | WAM | — | 49.2 |
| DiT4DiT | WAM | — | 50.8 |
| LDA-1B | WAM | — | 55.4 |
| Fast-WAM† | WAM | 56.0 | 76.3 |
| Joint-WAM† | WAM | 64.1 | 78.8 |
| LAWA (Ours) | WAM | 65.6 | 80.8 |
LAWA exceeds the matched Fast-WAM by 9.6 points in few-shot and 4.5 points with full data.
LIBERO-Plus zero-shot robustness
Micro-average success across seven perturbation types.
| Method | Language | Camera | Lighting | Motion | Noise | Texture | Layout | Total |
|---|---|---|---|---|---|---|---|---|
| OpenVLA | 0.8 | 3.5 | 23.0 | 8.1 | 34.8 | 15.2 | 28.5 | 15.6 |
| OpenVLA-OFT | 56.4 | 31.9 | 79.5 | 88.7 | 93.3 | 75.8 | 74.2 | 69.6 |
| UniVLA | 1.8 | 46.2 | 69.6 | 69.0 | 81.0 | 21.2 | 31.9 | 42.9 |
| WorldVLA | 0.1 | 27.9 | 41.6 | 43.7 | 17.1 | 10.9 | 38.0 | 25.0 |
| π0 | 13.8 | 6.0 | 58.8 | 85.0 | 81.4 | 79.0 | 68.9 | 53.6 |
| π0-FAST | 65.1 | 21.6 | 61.0 | 73.2 | 73.2 | 74.4 | 68.8 | 61.6 |
| Fast-WAM† | 24.9 | 50.7 | 76.9 | 89.2 | 62.3 | 58.0 | 67.7 | 60.0 |
| Joint-WAM† | 47.2 | 65.3 | 91.8 | 94.8 | 57.6 | 61.0 | 78.9 | 70.4 |
| LAWA (Ours) | 69.2 | 66.7 | 62.8 | 96.2 | 64.5 | 85.5 | 78.6 | 74.4 |
End-to-end latency
Milliseconds per action chunk on one NVIDIA A800 GPU.
LAWA is 42.9% faster than Joint-WAM at comparable success, while substantially outperforming the faster Fast-WAM.
Why the latent sequence matters
RoboCasa full-data success under inference-time perturbations.
| Latent input | Success rate |
|---|---|
| Gaussian noise | 52.2 |
| Temporal shuffle | 56.4 |
| Unperturbed | 80.8 |
Scales with action-free video
From 10% to 100% of the video corpus, LAWA gains 3.6 points in full-data and 4.0 points in few-shot training.
Focuses on interaction regions
In the illustrative rollout, LAWA follows the manipulated object and task-relevant region rather than diffuse background appearance.
Real-world tasks
From fine-grained assembly to long-horizon manipulation.
Four tasks test complementary requirements on a UFACTORY xArm7: Gear and Battery emphasize assembly precision; Block and Laboratory require multi-stage execution.
40.0%LAWA average with 25% of demonstrations
67.5%LAWA average with full demonstrations
31.2point gain over Fast-WAM at 25% data
33.8point gain over Fast-WAM at full data
With only 25% of demonstrations—50 trajectories per task—LAWA reaches 40.0% average success, surpassing the 33.8% achieved by Fast-WAM with the full training set. In the full-data setting, LAWA improves by 45 points on both long-horizon tasks.
Robot rollouts
LAWA rollouts in the real world.
Each task includes two representative LAWA rollout recordings captured from different angles. Every video is encoded at 4× its original speed.
Gear assembly
A fine-grained task requiring the robot to position the gear so its teeth mesh with the adjacent gear.
Battery insertion
A fine-grained assembly task in which the battery must be aligned and pressed until fully seated in the slot.
Block manipulation
A long-horizon task that requires multiple object transfers and sustained execution across stages.
Laboratory manipulation
A long-horizon sequence involving multiple containers and transfers across the workspace.
Citation
BibTeX
Citation metadata will be added after the anonymous submission stage.
BibTeX forthcoming.