WorldEvals¶

The Inspect Evals for robotics.

A curated catalog of physical-AI / VLA benchmarks built on Inspect Robots. Each benchmark lives in its own repository (so it owns its release cadence, dependencies, and hardware notes); WorldEvals is the index that ties them together.

GitHub Add a benchmark

Two views of the collection

inspect-robots list tells you what Inspect Robots tasks are installed. worldevals list tells you what benchmarks exist and how to get them.

Benchmarks¶

4 benchmarks · 26 tasks

StationeryBench

5 bimanual desk-stationery tasks, one fixed setup each: uncap a marker, open a box, take out an eraser, and close the lid, pull the middle pad from a sticky-note stack, pour paper clips into a lifted bowl, and hand a ruler between arms cup to cup.

5 tasks · bimanual · stationery bimanual manipulation

Repo · alpha
```
pip install "stationerybench @ git+https://github.com/robocurve/stationerybench"
```
RoboDojo-RC Tier 1

6 of RoboDojo's real-world manipulation tasks on bimanual YAM arms, with the pinned rubrics and the video judge behind Robocurve's published scores: pack and pour fruit, stack bowls, store objects in a safe, cap a pen, stand up bottles, and sort objects into baskets. Each trial scores a rubric level from 0 to 1.

6 tasks · bimanual · robodojo bimanual manipulation llm-agents

Repo · alpha
```
pip install "robodojo-rc @ git+https://github.com/robocurve/robodojo-rc"
```
RoboHarm

5 fixed-scene tasks with harmful instructions, with the label rubric and the video-and-transcript judge behind Robocurve's published RoboHarm results: a knife next to a doll, an aerosol can and a burner, a screwdriver and a toaster, a power bank and water, and two cleaning-product containers. Each trial gets one label, such as a safety refusal or a harmful action.

5 tasks · bimanual · safety refusal bimanual manipulation llm-agents

Repo · alpha
```
pip install "roboharm @ git+https://github.com/robocurve/roboharm"
```
DataCenterBench Tier 1

10 steps from data-center maintenance procedures as bimanual tasks, each one fixed setup on a board, with the rubrics and the two-judge photo scoring behind Robocurve's published scores: unplug and plug a charger, turn a knob, flip a switch, turn a key, open a cover, move a slider, enter a keypad code, join an Ethernet plug, and press a cable into a clip. Each trial scores 0, 0.5 or 1.

10 tasks · bimanual · data-center bimanual manipulation llm-agents

Repo · alpha
```
pip install "datacenterbench @ git+https://github.com/robocurve/datacenterbench"
```

Browse from the command line¶

pip install worldevals

worldevals list                 # all benchmarks
worldevals list --tag bimanual  # filter by tag
worldevals info stationerybench # repo, install command, task keys
worldevals tasks                # Inspect Robots tasks installed locally, by benchmark

For LLMs: llms.txt · llms-full.txt.