Environments

Registry

RoboSmith uses an environment registry as a searchable catalog. The default file is configs/env_registry.yaml and currently contains 30 entries across classic control, MuJoCo, Gymnasium-Robotics, and Isaac Lab.

Each entry becomes an EnvEntry:

from robosmith.envs.registry import EnvRegistry

registry = EnvRegistry()
entry = registry.get("mujoco-ant")
print(entry.env_id)       # Ant-v5
print(entry.framework)    # gymnasium
print(entry.task_tags)    # locomotion, walk, run, quadruped, forward

Searching

matches = registry.search(
    robot_type="arm",
    env_type="tabletop",
    framework="gymnasium",
    tags=["pick", "place"],
)

Filters are combined. Tags are scored and sorted by match count. CLI filters use case-insensitive substring matching:

robosmith envs
robosmith envs --framework gym
robosmith envs --robot arm
robosmith envs --env-type tabletop
robosmith envs --tags "pick place"

Task Matching

Environment synthesis uses match_task_to_env():

from robosmith.config import TaskSpec
from robosmith.stages.env_synthesis.synthesis import match_task_to_env

spec = TaskSpec(
    task_description="A quadruped runs forward",
    robot_type="quadruped",
    environment_type="floor",
)
match = match_task_to_env(spec, registry)

print(match.entry.id)
print(match.score)
print(match.match_reason)

Matching order:

1. Use TaskSpec.environment_id directly when present.
2. Search by robot type, environment type, framework, robot model, and extracted task tags.
3. Relax the environment type constraint.
4. Fall back to robot type plus tags.
5. Reject weak matches with no tag overlap.

Creating Environments

make_env() routes an EnvEntry through EnvAdapterRegistry.

from robosmith.envs.wrapper import make_env

env = make_env(entry, render_mode="rgb_array", max_episode_steps=500)
obs, info = env.reset()
obs, reward, terminated, truncated, info = env.step(env.action_space.sample())
env.close()

The wrapper builds an EnvConfig from keyword arguments and passes it to the framework-specific adapter.

Adapters

Adapter	Frameworks	Dependency status
GymnasiumAdapter	gymnasium, gym	Available with .[sim].
IsaacLabAdapter	isaac_lab	Requires Isaac Lab and Isaac Sim setup.
LIBEROAdapter	libero	Requires LIBERO installation.
ManiSkillAdapter	maniskill	Requires .[maniskill] and simulator dependencies.
CustomMJCFAdapter	custom_mjcf	For custom MJCF or URDF-like assets.

Inspect adapter availability:

from robosmith.envs.adapter_registry import EnvAdapterRegistry

adapters = EnvAdapterRegistry()
print(adapters.list_available())
print(adapters.list_all())

EnvConfig

from pathlib import Path
from robosmith.envs.adapters import EnvConfig

cfg = EnvConfig(
    render_mode="rgb_array",
    seed=42,
    max_episode_steps=500,
    num_envs=16,
    asset_path=Path("robot.xml"),
    extra={"camera_name": "front"},
)

Framework-specific options belong in extra.

Custom Registry

Use a custom registry file with ForgeConfig.env_registry_path or robosmith.yaml.

env_registry_path: ./my_envs.yaml

Example entry:

environments:
  - id: my-franka-reach
    name: My Franka Reach
    framework: gymnasium
    env_id: MyFrankaReach-v0
    robot_type: arm
    robot_model: franka
    env_type: tabletop
    task_tags: [reach, franka, arm, target]
    obs_type: state
    action_type: continuous
    description: Franka reaches to a custom target.
    source: local