How to fix locations of objects in between environment resets?

[ameesh-shah]

Hi,

In order to further simplify the tasks provided in Safety Gymnasium, I'd like to fix the locations of the objects that are sampled upon each reset. I looked into the source code, but it seems that there is no easy way to do this without either building a new placements dict (which are generated in full during running) or changing the Underlying object source code to not repeatedly sample a new layout upon each reset. Is there a way to fix the locations of objects so that they do not change after resetting the environment? Thanks in advance!

[muchvo]

Please refer to the provided code. We've organized several boolean flags using dataclasses. You can find the relevant classes in the code, particularly this one:

class MechanismConf:
    r"""Mechanism options.

    Starting position distribution.

    Attributes:
        randomize_layout (bool): If false, set the random seed before layout to constant.
        continue_goal (bool): If true, draw a new goal after achievement.
        terminate_resample_failure (bool): If true, end episode when resampling fails,
        otherwise, raise a python exception.
    """

    randomize_layout: bool = True
    continue_goal: bool = True
    terminate_resample_failure: bool = True

This configuration is an attribute self.mechanism_conf within the base class BaseTask. Therefore, you should add the line self.mechanism_conf.randomize_layout = False in the __init__ method of the specific task class you're working on.

[ameesh-shah]

Thanks very much for your help!

[ameesh-shah]

Hi, sorry, one more question - I'd like to enforce certain boundaries for the agent (so that it can't go beyond a certain point in the map.) I noticed that the wall objects are specific to certain tasks that I'm not using, and that the .pos attribute of the agent is not mutable. How would I be able to enforce these boundaries on the map that the agent cannot go past?

[muchvo]

If you're looking to add walls to an environment, please refer to the implementation in this class. Note that the walls implemented by this class are penetrable by default. To make them solid, you should remove the following lines:

'contype': 0,
'conaffinity': 0,

By doing so, the walls will have a physical presence. You'll just need to reimplement a class according to your needs and integrate it into your specific task like so:

self._add_geoms(Hazards(num=8, keepout=0.18))

Since walls are immovable objects, they fall under the 'geom' category. For more information on the categorization of objects, please refer to the documentation here.

If you wish to directly change the position of an object, you would need to use MuJoCo's interface. You can find several functions that accomplish this here:

    def body_com(self, name):
        # Get the center of mass of a named body in the simulator world reference frame.
        return self.data.body(name).subtree_com.copy()

    def body_pos(self, name):
        # Get the position of a named body in the simulator world reference frame.
        return self.data.body(name).xpos.copy()

    def body_mat(self, name):
        # Get the rotation matrix of a named body in the simulator world reference frame.
        return self.data.body(name).xmat.copy().reshape(3, -1)

    def body_vel(self, name):
        # Get the velocity of a named body in the simulator world reference frame.
        return get_body_xvelp(self.model, self.data, name).copy()
...
    def set_state(self, value):
        """
        Sets the state from a dictionary.

        Args:
            value (dict): The desired state.
            call_forward: Optionally call sim.forward(). This is called by default if
                the udd_callback is set.
        """
        self.data.time = value['time']
        self.data.qpos[:] = np.copy(value['qpos'])
        self.data.qvel[:] = np.copy(value['qvel'])
        if self.model.na != 0:
            self.data.act[:] = np.copy(value['act'])

Feel free to reach out if you have any further questions or require additional assistance.