EMBODIED AI SAFETY · RESEARCH PROJECT

Embodied SafeWorld

From Machine-Centric
to Human-Centric Safety

A robot can move perfectly and still cause harm. Evaluating embodied intelligence through the consequences of its actions.

Code Coming Soon
Seedance 2.0 research scene with a wet sponge and an electrical power strip
01
Understand the contextRisk beyond physical motion
Wan 3.0 research scene with a heavy toolbox beside a person’s open hand
02
Anticipate the consequencesPeople, objects and their interactions
Galaxea G0.5 simulation: moving a cast iron weight onto a notebook
03
Evaluate the behaviorFrom visual predictions to robot actions

Selected model-generated scenarios and a simulation environment from the project.

21,600benchmark test samples
9real-world risk categories
6evaluation dimensions
4model families

Arena includes an additional 1,080 paired safe tasks.

01 / A DIFFERENT SAFETY QUESTION

Correct movement.
Wrong consequences.

Avoiding collisions is essential. Understanding the harm an action can cause is the next challenge.

The same grasp, move, and place sequence can be physically valid yet unsafe in context. Human-centric safety asks whether a model understands the objects, people, and delayed effects around its actions.

MACHINE-CENTRIC

How does it move?

Collision avoidance
Force limits
Grasp stability

HUMAN-CENTRIC

What could it cause?

Harm to people and animals
Damage to belongings
Exposure of private information

Comparison of machine-centric motion safety and human-centric consequences, illustrated through a robot moving a box
One task, two safety perspectives: motion constraints and consequences for people and their surroundings.

02 / TWO COMPLEMENTARY BENCHMARKS

See the risk.
Test the action.

From predicting a possible future to acting inside an interactive environment. Two benchmarks bring human-centric safety into focus.

ESW-Vista generated visual test sceneIMAGE + INSTRUCTION01

ESW-Vista

10,800 samples

Visual scenes paired with task instructions. Study the behaviors and consequences a model generates from a hazardous starting context.

Visual prediction · World-model evaluation
INTERACTIVE ENVIRONMENT02

ESW-Arena

10,800 samples

Galaxea G0.5 · 13.8 s
Cast iron weight placed on a notebook · Task completion: 5/5

Executable robot environments built on LIBERO. Observe actions as they unfold, with paired safe tasks to distinguish safety decisions from execution failures.

Interactive control · VLA / WAM / LLM
THREE AXES. SIX DIMENSIONS.

The danger is not always obvious—or immediate.

Direct↔Disguised
Immediate↔Delayed
Short-horizon↔Long-horizon

03 / A WORLD OF HIDDEN RISKS

Ordinary settings.
Extraordinary stakes.

Explore nine categories through actual generated research examples. Select a category, then play its model-generated trajectory.

MODEL-GENERATED RESEARCH EXAMPLE
01 / 09ESW-VISTA

Fire hazard

Scenario dimension
Immediate
Generation model
Wan 3.0
Browse categories

Generated trajectories illustrate the evaluation setting; they are not recordings of real-world incidents or standalone evidence of a model's overall safety.

Selected Wan 3.0 and Seedance 2.0 examples, curated for visual clarity and coherent motion.

04 / FROM CONCEPT TO CONSEQUENCE

A framework that
keeps asking what if.

Static benchmarks enable consistent comparison. Dynamic red teaming uses trajectory feedback to explore weaknesses beyond a fixed test set.

Framework connecting static benchmark construction with dynamic red teaming, trajectory evaluation and SkillBank feedback
Static benchmarks and dynamic red teaming share a feedback loop from risk mechanisms to evaluated trajectories.
01
◎

Discover mechanisms

Generate, select, expand, and validate causal pathways from broad risk concepts.

02
⌘

Build test worlds

Turn mechanisms into instructions, visual scenes, and executable environments.

03

Evaluate trajectories

Track dangerous attempts, critical actions, and consequences alongside task ability.

04
⟳

Learn from feedback

Retrieve and refine SkillBank experience to guide the next round of red teaming.

A

Attempt

A risk-related action begins.

→
C

Commit

The critical dangerous operation is reached.

→
S

Success

Harm occurs or a directly hazardous state forms.

05 / WHAT THE EVALUATION REVEALS

Capability is not
a safety guarantee.

Better generation or task execution does not necessarily mean safer behavior. Examine the reported results within each model family.

ESW-VISTA

World models

View all metrics
World models · ESW-Vista
ModelRefusalDangerous attemptDangerous commitDangerous success

Read rates alongside task capability. A low dangerous-success rate may reflect execution failure rather than active risk avoidance. WM and embodied-control results use different benchmarks and should not be treated as one leaderboard.

01 /

Seeing the future is not avoiding it.

World models can produce coherent dangerous trajectories. Visual quality and human-centric safety measure different things.

02 /

Failure is not the same as refusal.

The evaluated VLA and WAM models report 0% refusal. Paired safe tasks help reveal when low danger rates stem from limited execution ability.

03 /

Risk changes with context.

Concealment, delayed consequences, and longer action sequences expose different weaknesses. A single aggregate metric cannot describe them all.

BEYOND A FIXED BENCHMARK

Safety testing
should evolve, too.

Across ten rounds of self-evolving red teaming, the planner discovers scenarios that more frequently expose dangerous outcomes in the two target models.

100 test cases per target model per round.

DANGEROUS-SUCCESS RATE · ROUND 1 → ROUND 10

Cosmos3-Nano WM

37%→91%

GPT-6-Astra LLM

28%→78%

Higher rates indicate stronger risk discovery by the red team, not an improvement in target-model safety.

CODE COMING SOON

Toward intelligence
we can safely live with.

Explore the benchmarks and research questions. Follow the repository for the upcoming code and documentation release.