Configure Tasks and Scenes¶

Task configurations own episode timing, task assets, surrounding geometry, reset randomization, task observations, and success criteria. Select a registered configuration first, then inspect that task's config instead of assuming every task exposes the same fields.

Shared environment settings¶

Every maintained task derives from BaseEnvCfg in source/ambench/ambench/tasks/base/base_env_cfg.py.

Command-line tools may override values such as the environment count after loading the registered config. Record both the environment ID and launcher arguments when comparing runs.

Maintained task configs¶

These are current software defaults, not an immutable benchmark protocol. Read the selected class before overriding it, because derived registered configs may replace the robot profile or adjust task values.

Success and progress criteria¶

Final success produces terminated=True. Named criteria are rollout progress signals; the evaluator marks a criterion complete if it becomes true at least once.

Every maintained task calls the shared reset before restoring its own state. Articulated tasks restore their joints; object tasks restore or retain their configured/event-randomized poses as appropriate; attached-object tasks reattach the peg, ball, or sponge; and stateful tasks clear hold, grasp, contact, or visibility buffers.

Change a task value¶

For a maintained variant, define a config subclass next to the task rather than mutating a shared class-level object. This example changes only Press Button task settings:

from isaaclab.utils import configclass

from ambench.tasks.press_button.press_button_env_cfg import PressButtonEnvDefaultCfg


@configclass
class PressButtonLongEpisodeCfg(PressButtonEnvDefaultCfg):
    episode_length_s = 30
    button_pressed_threshold = 0.006

A new public variant also needs an explicit register_env(...) entry in the task package's __init__.py. Use Environment IDs for the naming contract and Add a Task for the registration and validation path.

Configure scene geometry and assets¶

Task configs use Isaac Lab asset configs such as ArticulationCfg, RigidObjectCfg, and scene helpers such as WallSceneCfg. Keep related derived values together: moving a wall may require recomputing the local offset or initial position of a button, hole, lever, slider, valve, or peg fixture.

Simple reusable surrounding geometry belongs in source/ambench/ambench/scenes/. Task-specific object counts, asset choices, poses, and success thresholds remain in the task config. Do not point a public config at a machine-local asset path.

Verify success and observations¶

Config fields only define the inputs to task behavior. The task's *_env.py remains authoritative for _get_observations() and _get_success(). After changing geometry, thresholds, or event ranges, run one bounded environment and exercise the smallest relevant scripted policy before scaling the number of environments.

See Physics, Disturbances, and Randomization for the current event surfaces and Actions, Observations, and Cameras for the environment-level policy contract.