Branch from one state.
Execute each policy proposal from the same saved state. Record the immediate visual change and the branch outcome under the collection protocol.
VISION · LANGUAGE · ACTION A PREDICTIVE SKILL FRAMEWORK
EmbodiedSkills↗
Before the action,
consider the consequence.
World Action Model-augmented Cascaded Skills
for Vision-Language-Action Agent
SCROLL TO DISCOVER
Recorded MetaWorld simulation.
Animated paths illustrate the method.
01 / THE IDEAPOLICY + PREDICTION
A policy proposes.
A world model
looks ahead.
Several actions can look plausible.
Their consequences can be very different.
EmbodiedSkills gives a frozen action policy a predictive decision layer. It anticipates the effects of candidate actions, compares their expected outcomes, and executes one original proposal.
Reusable skills connect perception, prediction and physical execution. The policy keeps its capabilities. The additional models learn how to choose.
Follow the decision loop ↓05
Cascaded stagesOne shared feedback loop
03
World action modelsIndependent consequence estimates
01
Original action selectedThe policy remains frozen
02 / THE METHODEXPLORE THE CASCADE
Five stages.
Always in the loop.
Observation becomes a proposal.
Prediction informs a choice.
Execution creates the next observation.
CURRENT OBSERVATION
Scene + instruction + state
FROZEN POLICY · LEARNED CONSEQUENCESFOLLOW THE INFORMATION
01 / OBSERVEMake the present
See the full architecture ↘Make the present
machine-readable.
Frozen visual and language encoders turn the current scene and instruction into a shared context, alongside observable state and elapsed progress.
- SKILLS
- Scene observation · Context encoding
- OUTPUT
- Visual tokens + goal embedding + state
03 / ANTICIPATING CONSEQUENCESTHREE VIEWS OF THE SAME FUTURE
Forecast the effect.
Compare the agreement.
Each candidate passes through three independent world action models. Selection accounts for both their estimated outcomes and how consistently they predict the effect.
POLICY PROPOSALS
A₀Direct
A₁
A₂
A₃
A₄
LATENT CONSEQUENCE MODELS
WAM 01Predict + evaluate
WAM 02Predict + evaluate
WAM 03Predict + evaluate
SHARED COMPARATORBeyond a
Beyond a
single forecast.
Outcome scores
Prediction disagreement
Predicted-effect agreement
SELECT AN ORIGINAL PROPOSAL A₃
Animated schematic · candidate A₃ shown for illustration
Sk = b̃k + 0.25 (p01k − p10k)
Latent, not pixel-space.
Predict compact visual changes without rendering candidate futures during deployment.
The direct action stays available.
Every alternative is compared against the frozen policy’s original proposal.
Shared visual encoding.
The members reuse pretrained features. Adaptation focuses on prediction and comparison.
04 / RECORDED EVIDENCEONE OBSERVATION. FIVE POSSIBILITIES.
A small difference.
A different outcome.
Explore real CLIPort candidate branches.
Switch the task, then inspect each proposal.
LANGUAGE GOAL
“Pack the hexagon in the brown box.”
THE SHARED STARTt
CANDIDATE A₃ · SELECTEDt + H
Images are immediate post-action observations. Outcome labels describe the complete recorded branch; they are shown here for analysis and are not inputs to the selector.
05 / THE FEEDBACK LOOPRECORDED ACTION REPLAY
Act. Observe.
Decide again.
The next decision begins where the last action left the robot. On MetaWorld, the selector receives a new observation after every executed chunk.
93low-level steps
≤ 8steps per decision
Frames replay the saved actions from successful episodes. Initial and final observations were checked against the recorded trajectories.
Dunk the basketball into the basketINITIAL OBSERVATION
STEP 000 / 093Every marker is a decision.A₀ = direct proposal
06 / EXPERIMENTSFROZEN POLICIES. PREDICTIVE DECISIONS.
Different backends.
A shared principle.
Continuous robot control, spatial manipulation
and long-horizon task chains.
COMPLETE-TASK SUCCESS
78.64%
+2.36 percentage points
50 manipulation tasks with a frozen π₀.₅ policy. The loop compares action chunks at each decision.
2,500 evaluation episodesDirect policyEmbodiedSkills
Same frozen policy backbone; success is measured on the complete task.
| Benchmark | Frozen policy | Direct | EmbodiedSkills | Gain |
|---|---|---|---|---|
| MetaWorld MT50 | π₀.₅ | 76.28% | 78.64% | +2.36 pp |
| CLIPort | CLIPort | 33.94% | 36.83% | +2.89 pp |
| CALVIN | FLOWER | 77.10% | 77.90% | +0.80 pp |
07 / LEARNING FROM BRANCHESOBSERVED TRANSITIONS → PREDICTIVE SKILLS
Learn the consequences.
Keep the policy.
Matched starting states provide comparable action outcomes. Learning is concentrated in the predictive and comparison modules.
Learn to anticipate.
Train three independent WAMs with observed latent transitions and action-quality supervision. The spatial model warms up dynamics before joint training.
Learn to compare.
Freeze the WAMs and train the shared comparison head on joint outcomes. It estimates the benefit and risk of replacing the direct proposal.
Action policy, visual backbone and language encoder stay frozen.
Future observations and branch outcomes supervise training.
08 / THE COMPLETE PICTUREOPEN THE FIGURES TO EXPLORE
One cascade.
Composable skills.
Defined inputs and outputs keep policy proposals, predicted effects and executed actions connected across backends.
09 / RESEARCH & RESOURCESEMBODIEDSKILLS
World Action Model-augmented
Cascaded Skills for
Vision-Language-Action Agent
Cite this work +
@article{wang2026embodiedskills,
title={EmbodiedSkills: World Action Model-augmented Cascaded Skills
for Vision-Language-Action Agent},
author={Wang, Wei and Zhang, Wenqiao and Lin, Yutong and Yuan, Yuqian
and Lin, Tianwei and Mao, Jinhao and Fan, Zhenxuan
and Gao, Mingjian and Dai, Yang and Li, Wentong and Lv, Zheqi
and Dong, Zheng and Niu, Yingjie and Zhu, Jiaqi and Xiao, Jun
and Li, Chao and Zhuang, Yueting},
journal={arXiv preprint arXiv:2609.01281},
year={2026}
}