Unified Evaluation for Embodied Intelligence

Independent and comprehensive evaluation for the perception, generation, action, and physical agentic capabilities of embodied AI models

Evaluation methods

Embodied AnalysisAction Model Capability Index

Evaluation Method

gpt-6-astra 1388

claude-opus-5 1332

claude-fable-5-1 1226

gpt-5.6-sol 1057

claude-opus-4-8 1035

gemini-3.8-flash 1017

grok-4.6 1017

deepseek-v4.1-flash 967

qwen3.8-max-0902 960

glm-5.3-flash 948

π0.5 941

muse-spark-1.3 899

kimi-k3 874

gpt-5.5 869

inkling 470

15 modelsElo Average

Embodied AnalysisWorld Model Capability Index

Evaluation Method

Wan 3 1161

MiniMax H3 1129

Kling v3 Pro 1101

Seedance 2 1082

Gemini Omni Flash 953

Grok Imagine 1.5 932

Cosmos 3 Super 831

Veo 3.1 812

8 modelsElo Average

Action Model

Action Model Evaluation

Action models are evaluated on their capabilities to drive physical embodiments to complete complex long-horizon tasks.

Goal Completion

Does the embodiment complete the specified task within time limit?

Generalization

Does the model generalize across different tasks and embodiments?

Efficiency

Does the model produce a clean and efficient execution trajectory?

World Model

World Model Evaluation

World models are evaluated on multiple axes including physical plausibility, goal completion, and instruction following.

Physical Plausibility

Do gravity, contact, and object solidity remain plausible, without clipping or teleportation?

Goal Completion

Does the clip reach the goal state described by the task?

Instruction Following

Does the robot follow the instruction while keeping the same embodiment?