Scenarios
The available tasks, and how scenarios work.
A scenario is a robot, a scene and a task with success criteria. Verify runs your policy in a scenario for several seeded episodes; Studio lets you open the same scenarios, edit them, and save your own.
Available scenarios
| Scenario | Difficulty | Task | VLA variant |
|---|---|---|---|
messy_room | Easy | Reach and touch the red cube, ignoring a distractor | messy_room_vla |
navigation | Easy | Move the end effector to a target marker | |
object_sorting | Easy | Sort red and blue cubes into matching bins | |
mug_placement | Easy | Place a mug upright on a coaster | |
kitchen_pickup | Medium | Pick up a mug and place it on a plate (also _easy, _hard, _no_distractor) | kitchen_pickup_vla |
table_clearing | Medium | Clear the objects on a table into a bin | |
warehouse_retrieval | Medium | Mobile robot fetches a box from a shelf to a staging area | |
dual_arm_handover | Medium | One arm hands a cube to the other | dual_arm_handover_vla |
block_stacking | Hard | Stack red, blue and green blocks in order | block_stacking_vla |
precision_placement | Hard | Place colored cylinders on matching markers | |
obstacle_avoidance | Hard | Reach a target without touching obstacles |
The VLA variants are for Vision-Language-Action models, which take a language instruction and camera images instead of the low-dimensional state. See the scenario reference for the observation keys, action limits and success criteria of each one.
How a scenario runs
- The scenario's
scenario.yamlis compiled to MuJoCo MJCF. - Each episode resets the scene and calls your policy once per step with an observation (a dictionary of arrays and numbers).
- Your policy returns an action vector. It is applied to the robot's actuators and the physics advances.
- The episode runs until the step or time limit. Success criteria are checked against the final state.
- Contacts, forces, trajectory and video are recorded for grading.
Scoring
Each episode is scored on progress (how far the task got, 0 to 10) and safety (collisions, forces and workspace limits, 0 to 10). Across the episodes the job also gets a consistency score and a success rate, and a vision-language model reviews the video against the success criteria. See Verify for the result fields.
Make your own
Build a scenario in Studio or by editing a scenario.yaml, and save it to your
library. Scenarios you save can be submitted to Verify by id.