Enable passing custom logging services

CITATION.cff

@@ -3,7 +3,7 @@ title: "RSL-RL: A Learning Library for Robotics Research"
 message: "If you use this work, please cite the following paper."
 repository-code: "https://github.com/leggedrobotics/rsl_rl"
 license: BSD-3-Clause
-version: 5.3.0
+version: 5.4.0
 type: software
 authors:
   - family-names: Schwarke

CONTRIBUTORS.md

@@ -35,6 +35,7 @@ Please keep the lists sorted alphabetically.
 * Fabian Jenelten
 * Lorenzo Terenzi
 * Marko Bjelonic
+* Markus Portugall
 * Matthijs van der Boon
 * Özhan Özen
 * Pascal Roth

docs/api/utils.rst

@@ -9,29 +9,34 @@ Logger
    :undoc-members:
 
 
-Utils
------
+Log Writer
+----------
 
-.. automodule:: rsl_rl.utils.utils
+.. automodule:: rsl_rl.utils.log_writer
    :members:
    :undoc-members:
 
 
-Wandb Utils
------------
+Wandb Log Writer
+----------------
 
-.. automodule:: rsl_rl.utils.wandb_utils
+.. automodule:: rsl_rl.utils.wandb_log_writer
    :members:
    :undoc-members:
 
 
-Neptune Utils
--------------
+Neptune Log Writer
+------------------
 
-.. automodule:: rsl_rl.utils.neptune_utils
+.. automodule:: rsl_rl.utils.neptune_log_writer
    :members:
    :undoc-members:
 
 
+Utils
+-----
 
+.. automodule:: rsl_rl.utils.utils
+   :members:
+   :undoc-members:
 

docs/guide/configuration.rst

@@ -34,8 +34,10 @@ Runner Configuration
 
 Currently, RSL-RL implements two runner classes:
 :class:`~rsl_rl.runners.on_policy_runner.OnPolicyRunner` and
-:class:`~rsl_rl.runners.distillation_runner.DistillationRunner`. The 
-:class:`~rsl_rl.runners.on_policy_runner.OnPolicyRunner` is configured as follows:
+:class:`~rsl_rl.runners.distillation_runner.DistillationRunner`, which are configured as follows.
+
+OnPolicyRunner
+^^^^^^^^^^^^^^
 
 .. list-table::
    :header-rows: 1
@@ -45,31 +47,19 @@ Currently, RSL-RL implements two runner classes:
      - Type
      - Default
      - Description
-   * - ``num_steps_per_env``
-     - int
-     - required
-     - Number of environment steps collected per iteration.
    * - ``obs_groups``
      - dict[str, list[str]]
      - required
      - Mapping from observation sets to observation groups coming from the environment. See :ref:`here
        <observation-configuration>` for more details.
+   * - ``num_steps_per_env``
+     - int
+     - required
+     - Number of environment steps collected per iteration.
    * - ``save_interval``
      - int
      - required
      - Number of iterations between checkpoints.
-   * - ``logger``
-     - str
-     - ``"tensorboard"``
-     - Logging service to use. Valid values: ``"tensorboard"``, ``"wandb"``, ``"neptune"``.
-   * - ``wandb_project``
-     - str
-     - required for W&B
-     - W&B project name used by the W&B writer.
-   * - ``neptune_project``
-     - str
-     - required for Neptune
-     - Neptune project name used by the Neptune writer.
    * - ``run_name``
      - str
      - missing
@@ -83,6 +73,19 @@ Currently, RSL-RL implements two runner classes:
      - ``None``
      - Compile mode for the PyTorch models to accelerate training.
        Valid values: ``None``, ``"default"``, ``"max-autotune-no-cudagraphs"``.
+   * - ``logger``
+     - str | dict
+     - ``"tensorboard"``
+     - Logging writer configuration. The plain strings ``"wandb"`` and ``"neptune"`` are 
+       still accepted but deprecated.
+   * - ``wandb_project``
+     - str
+     - --
+     - Deprecated. Pass ``project_name`` inside the ``logger`` configuration instead.
+   * - ``neptune_project``
+     - str
+     - --
+     - Deprecated. Pass ``project_name`` inside the ``logger`` configuration instead.
    * - ``algorithm``
      - dict
      - required
@@ -96,6 +99,9 @@ Currently, RSL-RL implements two runner classes:
      - required
      - Critic model configuration.
 
+DistillationRunner
+^^^^^^^^^^^^^^^^^^
+
 For the :class:`~rsl_rl.runners.distillation_runner.DistillationRunner`, the ``actor`` and ``critic`` keys are simply
 replaced by ``student`` and ``teacher`` keys, respectively:
 
@@ -120,6 +126,31 @@ replaced by ``student`` and ``teacher`` keys, respectively:
      - required
      - Teacher model configuration.
 
+Logger
+^^^^^^
+The ``logger`` key of the runner configuration defines the logging writer used to log training metrics and other
+information during training. RSL-RL supports TensorBoard, Weights & Biases, and Neptune out of the box. While
+TensorBoard does not require any configuration and is set using the plain string ``"tensorboard"``, the other logging
+backends are configured by passing a dictionary with the following keys:
+
+.. list-table::
+   :header-rows: 1
+   :class: no-wrap-type-column
+
+   * - Key
+     - Type
+     - Default
+     - Description
+   * - ``class_name``
+     - str
+     - required
+     - Logger class name. Valid values: ``"WandbLogWriter"``, ``"NeptuneLogWriter"``.
+   * - ``project_name``
+     - str
+     - required
+     - Name of the project.
+
+
 Algorithm Configuration
 -----------------------
 

docs/guide/overview.rst

@@ -14,6 +14,11 @@ Library Features
 RSL-RL is intentionally kept minimal and focuses on a small set of components that cover common robotics workflows while
 remaining easy to adapt. The following sections summarize the main features currently available.
 
+.. note::
+   Adding new algorithms, models, or loggers is straightforward and does not require modifying the library itself.
+   Custom classes can simply be passed as part of the configuration, enabling users to work with the pip version of the 
+   library.
+
 Algorithms
 ^^^^^^^^^^
 
@@ -198,7 +203,9 @@ not constrain the way an **Extension** may be implemented, allowing for arbitrar
 Utils
 ^^^^^
 **Utils** include various helpers for the library, such as a :class:`~rsl_rl.utils.logger.Logger` to record the learning
-process, or functions to resolve configuration settings.
+process, or functions to resolve configuration settings. For example, the :func:`~rsl_rl.utils.utils.resolve_callable` 
+function allows users to pass classes via the configuration dictionary, enabling the use of custom models, loggers, etc. 
+without modifying the library.
 
 .. _example-integration:
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "rsl-rl-lib"
-version = "5.3.0"
+version = "5.4.0"
 keywords = ["reinforcement-learning", "robotics"]
 maintainers = [
   { name="Clemens Schwarke", email="cschwarke@ethz.ch" },

rsl_rl/utils/__init__.py

@@ -5,6 +5,8 @@
 
 """Helper functions."""
 
+from .log_writer import LogWriter
+from .neptune_log_writer import NeptuneLogWriter
 from .utils import (
     check_nan,
     compile_model,
@@ -16,8 +18,12 @@
     split_and_pad_trajectories,
     unpad_trajectories,
 )
+from .wandb_log_writer import WandbLogWriter
 
 __all__ = [
+    "LogWriter",
+    "NeptuneLogWriter",
+    "WandbLogWriter",
     "check_nan",
     "compile_model",
     "get_param",

rsl_rl/utils/log_writer.py

@@ -0,0 +1,43 @@
+# Copyright (c) 2021-2026, ETH Zurich and NVIDIA CORPORATION
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+from __future__ import annotations
+
+import pathlib
+from abc import ABC, abstractmethod
+
+
+class LogWriter(ABC):
+    """Abstract base class for logging backends.
+
+    Log writers are configured via ``cfg["logger"]``, a dict with a ``"class_name"`` key pointing to the subclass and
+    any additional keys forwarded as constructor kwargs. The class is resolved via
+    :func:`~rsl_rl.utils.resolve_callable`. Only :meth:`add_scalar` must be implemented; all other methods are no-ops.
+    """
+
+    @abstractmethod
+    def add_scalar(self, tag: str, scalar_value: float, global_step: int) -> None:
+        """Log a scalar metric.
+
+        Args:
+            tag: Name of the metric.
+            scalar_value: Value of the metric.
+            global_step: Current training iteration.
+        """
+
+    def store_config(self, env_cfg: dict | object, train_cfg: dict) -> None:
+        """Upload environment and training configuration. Called once at training start."""
+
+    def save_model(self, model_path: str, it: int) -> None:
+        """Upload a model checkpoint."""
+
+    def save_file(self, path: str) -> None:
+        """Upload an arbitrary file."""
+
+    def save_video(self, video: pathlib.Path, it: int) -> None:
+        """Upload a video file."""
+
+    def stop(self) -> None:
+        """Finalize and close the logging run."""