ECCV 2026

EvoWorld

Imagine. Act. Evolve.

A World-Model-Centric Framework for Continuous Self-Evolution of Modular Embodied Skills

Boshi Zhang*1, Sen Cui*‡1, BaoHuaYin3, Youyi Kou4, Junyu Wu5, Zuo Pu1, Tao Xue1, Zhikang Chen2, Shanshan Wei1, Min Zhang6, Miao Liu1, Changshui Zhang†1, Zhang Tao†1

1Tsinghua University · 2University of Oxford · 3University of Sussex · 4Beijing Jiaotong University · 5Xiamen University · 6East China Normal University

* Equal contribution · ‡ Project leader · † Corresponding authors

The research film · 1:06 · 1080p · no audio, captions availableOpen video ↗
25.9→44.8%
Mean policy success
Four VLA backbones, five skill families · Table 1
29→70%
Real-robot success
Franka arm, 5 tasks × 20 trials · rebuttal study
4
Plug-in VLA experts
GR00T-N1.6 · π₀.₅ · π₀ · π₀-FAST

Abstract

Share the dynamics. Specialize the skills.

Current progress in physical intelligence largely relies on scaling monolithic Vision-Language-Action (VLA) models, yet real-world policy data remain fragmented across scenes and tasks. This mismatch limits transfer, exacerbates catastrophic forgetting, and impedes continual improvement. A modular design that shares dynamics while specializing skills is therefore a promising paradigm.

We introduce EvoWorld (EvoW), a world-model-centric framework for skill orchestration and iterative self-evolution. VLAs form an expandable library of pluggable experts. A high-level router selects experts conditioned on scene and task, while an action-conditioned video world model serves as the cognitive core for rollout-based planning. A vision-language evaluator delivers semantic scoring and diagnostic tags, enabling targeted updates to the world model, router memory, or specific experts rather than global retraining.

Vision-Language-Action modelsWorld modelsExpert routingRobot manipulationContinual skill refinement

Method

One loop. Four moves.

Every task passes through the same closed loop. Select a step to see what happens there.

Shared world model 01 02 03 04 GROUND ROUTE IMAGINE EVOLVE
01 · Ground

Start from a real scene.

A task description is matched against an offline repository of real-robot recordings with CLIP embeddings, and a grounded initial state is sampled, so every generated task begins in a scene the robot has actually seen.

Output: a grounded state o₀ for the task ℓ.

EvoWorld pipeline: task generation, contextual scene anchoring, strategic routing and imaginative simulation, VLM diagnosis and recursive evolution
The full pipeline: (1) task generation, (2) contextual scene anchoring, (3) strategic routing and imaginative simulation, (4) VLM diagnosis and recursive evolution. Select to open full resolution.

Demos

Imagine first. Then act.

Generated world-model rollouts and physical robot footage are shown separately and labeled.

View 1View 2Wrist
S1 · Pick & place

Pick & place

“Put the yellow block in the blue cup.”

World-model rollout · 1× · generated video

View 1View 2Wrist
S3 · Reorientation

Reorientation

“Put the cup upright.”

World-model rollout · 1× · generated video

View 1View 2Wrist
S4 · Deformable

Deformable manipulation

“Use the white towel to wipe the table.”

World-model rollout · 1× · generated video

Results

Better skills, every backbone.

Mean success rises by 18.9 percentage points across four VLA backbones, with relative gains from +62.0% to +82.8%.

Per-skill success

Table 1 · five skill families · choose a backbone

BaseWith EvoWorld

Average across skills

All four backbones

25.9% → 44.8%

Mean over the four backbones

Full benchmark tables & additional figures
Policy success rates across five representative skill families and four VLA backbones. EvoW* denotes our method.
Backbone Setting S1 S2 S3 S4 S5 Avg Gain
GR00T-N1.6 Base 0.38 0.24 0.18 0.28 0.20 0.256 --
EvoW* 0.48 0.50 0.46 0.42 0.48 0.468 +82.8%
PI-0.5 Base 0.32 0.14 0.24 0.18 0.28 0.232 --
EvoW* 0.48 0.28 0.38 0.44 0.48 0.412 +77.6%
PI-0 Base 0.40 0.22 0.26 0.32 0.22 0.284 --
EvoW* 0.62 0.38 0.40 0.40 0.50 0.460 +62.0%
PI-0-fast Base 0.38 0.18 0.26 0.24 0.26 0.264 --
EvoW* 0.66 0.38 0.44 0.36 0.42 0.452 +71.2%
Task and scene diversity under executability constraints. Lower is better for similarity metrics; bold marks the best value in each row.
Metric EvoW RoboGen Behavior-100 RLBench MetaWorld ManiSkill2 GenSim
Number of tasks 100 106 100 106 50 20 70
Task description Self-BLEU 0.281 0.284 0.299 0.317 0.322 0.674 0.378
Task description Sentence-BERT similarity 0.312 0.165 0.210 0.200 0.263 0.194 0.288
Scene image ViT similarity 0.337 0.193 0.389 0.375 0.517 0.332 0.717
Scene image CLIP similarity 0.754 0.762 0.833 0.864 0.867 0.828 0.932

Citation

Build on this work.

@inproceedings{zhang2026evoworld,
  title={EvoWorld: A World-Model-Centric Framework for Continuous Self-Evolution of Modular Embodied Skills},
  author={Boshi Zhang and Sen Cui and {BaoHuaYin} and Youyi Kou and Junyu Wu and Zuo Pu and Tao Xue and Zhikang Chen and Shanshan Wei and Min Zhang and Miao Liu and Changshui Zhang and {Zhang Tao}},
  booktitle={European Conference on Computer Vision},
  year={2026}
}