Source code for torchrl.envs.custom.mujoco._humanoid_primitives
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
"""Humanoid macro actions used by MuJoCo examples."""
from __future__ import annotations
import torch
from torchrl.envs.transforms._primitive import TargetMacroAction
__all__ = ["HumanoidMacroAction"]
[docs]
class HumanoidMacroAction(TargetMacroAction):
r"""Structured action for humanoid actuator-control macros.
``HumanoidMacroAction`` is the thinnest possible specialization of
:class:`~torchrl.envs.transforms.TargetMacroAction`: the humanoid demo does
not solve a Cartesian inverse-kinematics problem, it asks the base env to
interpolate toward a low-level actuator-control destination. A plain
:class:`~torchrl.envs.transforms.MacroPrimitiveTransform` expands the target
into a fixed-length action sequence; the named factory below only documents
the intent and validates the target shape.
Examples:
>>> import torch
>>> from torchrl.envs import HumanoidMacroAction
>>> action = HumanoidMacroAction.reach_control(torch.zeros(1, 4), steps=2)
>>> action.target.shape
torch.Size([1, 4])
"""
[docs]
@classmethod
def reach_control(
cls,
target: torch.Tensor,
*,
steps: int = 16,
settle_steps: int = 0,
) -> HumanoidMacroAction:
"""Interpolate toward a low-level actuator-control target."""
return cls.move(target, steps=steps, settle_steps=settle_steps)
