Version 1.0

[jgillick]

This is a major version that cleans up some syntax and removes deprecated items. See the UPGRADE.md file for the full instructions on upgrading your environments.

Breaking changes

• Requires Genesis World 1.3.0+
• Removed deprecated params
• Refactored MDP functions and params.
• Changed *_attr params to point to the actual instances (e.g. robot_attr="robot" to robot=self.robot)
• NOTE: Fixed a reset bug to clear previous step values (observations, terminations, actions, etc). Your environment might need to enable observation normalization (in rsl-rl: obs_normalization=True) if it stops reaching convergence.

Other Changes

• Migrated to mono-repo
• Added genesis-forge-runtime package for running trained models on robot.
• Add VelocityActionManager
• Added a Pose2dCommand for commanding a target position and heading
• Added imu_projected_gravity observation to create estimated projected gravity from the IMU sensor.
• VelocityCommandManager: added angular velocity debug indicator
  
• Enhanced velocity tracking rewards (command_tracking_lin_vel/command_tracking_ang_vel): sensitivity can now be automatically derived from the command range
• Examples: Added a couple of wheeled robot examples, one including obstacle avoidance
• ContactManager: Added support for multiple target entities. (with_entities=list(...))
• Action managers: you can now drive several joints from one action with action_groups.
• Action managers: added action delay randomization (pass a min/max tuple instead of a fixed int)
• Fix ObservationManager: noise is applied before scaling
• Added higher-resolution debug arrows.
• Added observation clipping
• Fix action delay in setup/reset
• Fix envs_idx to be a tensor type on all managers and mdp reset functions
• Fix class name type: SkrlEnvWapper -> SkrlEnvWrapper
• Added unit test coverage
• Fixes for Genesis 1.4

MDP Functions (rewards, terminations, observations, entity resets)

This is the biggest change in this version. We're deprecating the params dictionaries in favor of data classes. For example:

Before

RewardManager(
    self,
    cfg={
        "base_height_target": {
            "weight": -50.0,
            "fn": rewards.base_height,
            "params": {
                "target_height": 0.3,
                "entity_attr": "robot",
            },
        },
    }
);

After

RewardManager(
    self,
    cfg={
        "base_height_target": {
            "weight": -50.0,
            "fn": rewards.base_height(
                target_height=0.3, 
                entity_attr="robot"
            ),
        },
    }
);

Note that params are passed directly into base_height. This allows for easier type checking and IDE autocomplete. Before, you'd need to open the mdp function documentation to get the list of params to type into the params dict.

This also changes how you write custom class-based MDP functions. Now they look like this:

@dataclass(kw_only=True, eq=False)
class lin_vel_z_l2(MdpFn):
    entity_manager: EntityManager = None

    def __call__(self, env: GenesisEnv) -> torch.Tensor:
        linear_vel = self.entity_manager.get_linear_velocity()
        return torch.square(linear_vel[:, 2])