USTC logoLumos logo

NexBench

Real-World Benchmark for Embodied AI Manipulation

Why NexBench

A benchmark built for the real world.

From world-model evaluation to industrial deployment, NexBench turns evidence into action.

01

The First World Model-Driven Benchmark

Leveraging the world’s first evaluation methodology powered by non-physical interactive world models, we simulate infinite scenarios without the need for costly hardware prototypes. This breakthrough eliminates the “sim-to-real” gap before deployment.

02

Industrial & Real-World Readiness

Move beyond lab metrics. Our platform focuses on rigorous testing for complex industrial scenarios and unpredictable real-world conditions, ensuring your robots are truly ready for prime time.

03

Automated Insights & Reporting

Stop manually crunching data. Our system automatically generates comprehensive test reports, providing actionable insights into performance, safety, and reliability in minutes.

04

Closing the Embodied Industry Loop

We don't just test; we connect. By aligning evaluation directly with industrial needs, we create a closed-loop ecosystem for Embodied AI—from algorithm training to scalable production.

World-model simulator evaluation

Scale reproducible manipulation scenarios with our own world-model simulator.

The simulation environment is generated in-house by NexBench rather than a third-party simulation platform, so models can be evaluated at scale and reproducibly without occupying real hardware.

Explore the evaluation

Real-world evaluation

Reproducible real-world evaluation through Station Lite.

Approved tasks move from NexBench to NexCore and Station Lite. Status, reports, and evidence return as structured evaluation records.

Browse sample rollouts

NexBench RealEval

Standardized execution and reviewed evidence.

Evaluation Process

Station Lite defines the execution environment. NexCore manages task dispatch and status return.

Lumos Station Lite dual-arm evaluation system

RealEval station

Lumos Station LiteDual-arm manipulation platform
Lumos Station Lite supports standardized real-world task execution and evidence capture.

Leaderboard

Reviewed rankings and publication records.

Inspect ranking distance, benchmark versions, sources, and update dates.

Ranked table

Leaderboard table

Published records maintained by the NexBench content team.

RankModel / OrganizationModel sizeSuccess rate
No published results match this view.

World-model simulator evaluation

Policy achieves physical interaction and completes evaluation in the world-model simulator.

World-model simulator closed-loop rollout diagram

Simulation engine

World simulator

A learned environment for large-scale robotic rollout and evaluation.

World-model simulator footage of robotic arms manipulating objects

Generate.

Rollout.

Evaluate.

Simulate future interactions before real-world execution.

  • Hardware-free rollouts

    Execute policy rollouts inside the simulator without requiring physical robots.

  • Scalable evaluation

    Benchmark many trajectories in parallel for faster and broader evaluation.

  • Safe testing

    Probe failures, edge cases, and OOD scenarios with reduced real-world risk.

The robot moves inside a generated world

Reality

Recorded footage

Recreated

World-model generated

Same recorded actions / two visual results

Participants

Teams advancing real-world evaluation.

Organizations represented in the current reference leaderboard.