Add skrl as a second RL backend behind a backend abstraction

docs/cli/play-train.en.md

@@ -41,6 +41,29 @@ uv run genelab train TASK_ID --gpus 4 --num_envs 4096
 `--num_envs` is total across ranks and must divide evenly by `--gpus`. Use
 `--num_envs_per_gpu` for per-rank semantics.
 
+## RL backends
+
+The training backend is chosen automatically from the type of the task's agent
+config — no flag required:
+
+| Agent config | Backend | Algorithms |
+|---|---|---|
+| `RslRlOnPolicyRunnerCfg` | `rsl_rl` (default) | PPO |
+| `SkrlAgentCfg` | `skrl` | PPO, A2C, SAC, TD3, DDPG |
+
+The [skrl](https://skrl.readthedocs.io) backend is optional — install it with the
+`skrl` extra (`uv sync` already includes it in this checkout; downstream users run
+`pip install genelab[skrl]`). Pick the algorithm via `SkrlAgentCfg.algorithm`.
+
+skrl trains in environment **timesteps** rather than learning iterations, so
+`--max_iterations N` sets the skrl timestep budget for a skrl task. Multi-GPU
+(`--gpus`) is supported by the RSL-RL backend only.
+
+```bash
+# A task registered with a SkrlAgentCfg routes through the skrl backend.
+uv run genelab train GeneLab-Franka-Pick-And-Place-skrl-v0 --num_envs 2048 --max_iterations 12000
+```
+
 ## Config overrides
 
 Any unknown option after the task id is treated as a dotted config override:

docs/cli/play-train.zh.md

@@ -40,6 +40,28 @@ uv run genelab train TASK_ID --gpus 4 --num_envs 4096
 `--num_envs` 表示所有 rank 的总数，必须能被 `--gpus` 整除。每 rank 语义用
 `--num_envs_per_gpu`。
 
+## RL 后端
+
+训练后端由 task 的 agent 配置类型自动选择，无需任何 flag：
+
+| Agent 配置 | 后端 | 算法 |
+|---|---|---|
+| `RslRlOnPolicyRunnerCfg` | `rsl_rl`（默认） | PPO |
+| `SkrlAgentCfg` | `skrl` | PPO、A2C、SAC、TD3、DDPG |
+
+[skrl](https://skrl.readthedocs.io) 后端为可选项——通过 `skrl` extra 安装（本仓库
+`uv sync` 已包含；下游用户执行 `pip install genelab[skrl]`）。算法通过
+`SkrlAgentCfg.algorithm` 选择。
+
+skrl 以环境 **timestep**（而非 learning iteration）计量训练量，因此对 skrl task，
+`--max_iterations N` 设定的是 skrl 的 timestep 预算。多 GPU（`--gpus`）仅
+RSL-RL 后端支持。
+
+```bash
+# 注册为 SkrlAgentCfg 的 task 会走 skrl 后端。
+uv run genelab train GeneLab-Franka-Pick-And-Place-skrl-v0 --num_envs 2048 --max_iterations 12000
+```
+
 ## 配置 override
 
 task id 后的任何未知选项都会被当作 dotted config override：

docs/concepts/rl-runner.en.md

@@ -1,21 +1,36 @@
 # RL Runner
 
-`genelab.rl` connects registered tasks to RSL-RL. It is deliberately thin: GeneLab owns task
-resolution, config mutation, VecEnv adaptation, logs, profiling hooks, and distributed launch
-helpers; RSL-RL owns the learning algorithm.
+`genelab.rl` connects registered tasks to a pluggable RL **backend**. It is deliberately thin:
+GeneLab owns task resolution, config mutation, env construction, the bridge lifecycle, logs,
+profiling hooks, and distributed launch helpers; the backend owns the learning algorithm.
+
+## Backends
+
+`train_task` / `play_task` are backend-agnostic dispatchers. The backend is chosen from the type of
+`TaskCfg.agent`:
+
+| Agent config | Backend | Algorithms |
+|---|---|---|
+| `RslRlOnPolicyRunnerCfg` | `rsl_rl` (default) | PPO |
+| `SkrlAgentCfg` | `skrl` | PPO, A2C, SAC, TD3, DDPG |
+
+Backends live under `genelab.rl.backends` and register themselves by config type;
+`select_backend(agent_cfg)` resolves one. Adding another library means adding a `Backend`
+(`train` / `play`) plus its agent-config dataclass — no change to the dispatcher or CLI.
 
 ## Training flow
 
 ```text
 TASKS.get(task_id)
 └── TaskCfg.env + TaskCfg.agent
     └── ManagerBasedRlEnv
-        └── RslRlVecEnvWrapper
-            └── rsl_rl OnPolicyRunner.learn()
+        └── select_backend(agent_cfg).train(TrainContext)
+            ├── rsl_rl:  RslRlVecEnvWrapper  → OnPolicyRunner.learn()
+            └── skrl:    GenelabSkrlWrapper  → SequentialTrainer.train()
 ```
 
 The main process writes `params/env.json`, `params/agent.json`, TensorBoard events, profiler traces,
-and checkpoints.
+and checkpoints. RSL-RL logs under `logs/rsl_rl/`, skrl under `logs/skrl/`.
 
 ## Playback flow
 
@@ -25,7 +40,7 @@ and checkpoints.
 |---|---|
 | `zero` | Returns zero actions. |
 | `random` | Uniform random actions. |
-| `trained` | Loads a checkpoint and calls RSL-RL inference policy. |
+| `trained` | Loads a checkpoint and calls the backend's inference policy. |
 
 Playback exits cleanly when the Genesis viewer closes or when `max_steps` is reached.
 

docs/concepts/rl-runner.zh.md

@@ -1,19 +1,34 @@
 # RL runner
 
-`genelab.rl` 把已注册 task 接到 RSL-RL。它刻意保持很薄：GeneLab 负责 task 解析、配置修改、
-VecEnv 适配、日志、profiling hook 和分布式启动 helper；RSL-RL 负责学习算法。
+`genelab.rl` 把已注册 task 接到可插拔的 RL **后端**。它刻意保持很薄：GeneLab 负责 task 解析、
+配置修改、环境构建、bridge 生命周期、日志、profiling hook 和分布式启动 helper；后端负责学习算法。
+
+## 后端
+
+`train_task` / `play_task` 是与后端无关的分发器。后端由 `TaskCfg.agent` 的类型决定：
+
+| Agent 配置 | 后端 | 算法 |
+|---|---|---|
+| `RslRlOnPolicyRunnerCfg` | `rsl_rl`（默认） | PPO |
+| `SkrlAgentCfg` | `skrl` | PPO、A2C、SAC、TD3、DDPG |
+
+后端位于 `genelab.rl.backends`，按配置类型自行注册，`select_backend(agent_cfg)` 负责解析。
+接入另一个库只需新增一个 `Backend`（`train` / `play`）及其 agent 配置 dataclass——分发器和
+CLI 无需改动。
 
 ## 训练流程
 
 ```text
 TASKS.get(task_id)
 └── TaskCfg.env + TaskCfg.agent
     └── ManagerBasedRlEnv
-        └── RslRlVecEnvWrapper
-            └── rsl_rl OnPolicyRunner.learn()
+        └── select_backend(agent_cfg).train(TrainContext)
+            ├── rsl_rl:  RslRlVecEnvWrapper  → OnPolicyRunner.learn()
+            └── skrl:    GenelabSkrlWrapper  → SequentialTrainer.train()
 ```
 
-main process 写入 `params/env.json`、`params/agent.json`、TensorBoard event、profiler trace 和 checkpoint。
+main process 写入 `params/env.json`、`params/agent.json`、TensorBoard event、profiler trace 和
+checkpoint。RSL-RL 写到 `logs/rsl_rl/`，skrl 写到 `logs/skrl/`。
 
 ## 回放流程
 
@@ -23,7 +38,7 @@ main process 写入 `params/env.json`、`params/agent.json`、TensorBoard event
 |---|---|
 | `zero` | 返回零动作。 |
 | `random` | 均匀随机动作。 |
-| `trained` | 加载 checkpoint，并调用 RSL-RL inference policy。 |
+| `trained` | 加载 checkpoint，并调用后端的 inference policy。 |
 
 Genesis viewer 关闭或达到 `max_steps` 时，回放会干净退出。
 

examples/franka_pick_and_place/src/genelab_franka_pick_and_place/skrl_cfg.py

@@ -0,0 +1,34 @@
+"""skrl PPO agent config for the Franka pick-and-place task.
+
+Mirrors ``ppo_cfg.py`` (RSL-RL) but drives the skrl backend. skrl trains in
+environment timesteps rather than learning iterations: with ``--num-envs 2048``
+the default ``timesteps`` below is ~24M env-steps, comparable to the RSL-RL run
+(500 iterations x 24 steps).
+"""
+
+from genelab.rl import SkrlAgentCfg, SkrlExperimentCfg, SkrlModelCfg
+
+
+def franka_pick_and_place_skrl_cfg() -> SkrlAgentCfg:
+    return SkrlAgentCfg(
+        algorithm="PPO",
+        seed=42,
+        timesteps=12_000,
+        rollouts=24,
+        learning_rate=3.0e-4,
+        discount_factor=0.99,
+        learning_epochs=5,
+        mini_batches=4,
+        gae_lambda=0.95,
+        ratio_clip=0.2,
+        entropy_loss_scale=0.005,
+        grad_norm_clip=1.0,
+        models=SkrlModelCfg(
+            hidden_dims=(256, 256, 128),
+            activation="elu",
+        ),
+        experiment=SkrlExperimentCfg(
+            experiment_name="franka_pick_and_place_skrl",
+            logger="tensorboard",
+        ),
+    )

examples/franka_pick_and_place/src/genelab_franka_pick_and_place/tasks.py

@@ -19,9 +19,11 @@
     FrankaPickAndPlaceRobotCfg,
     get_franka_pick_and_place_robot_cfg,
 )
+from genelab_franka_pick_and_place.skrl_cfg import franka_pick_and_place_skrl_cfg
 
 TASK_ID = "GeneLab-Franka-Pick-And-Place-v0"
 CARTESIAN_TASK_ID = "GeneLab-Franka-Pick-And-Place-Cartesian-v0"
+SKRL_TASK_ID = "GeneLab-Franka-Pick-And-Place-skrl-v0"
 ROBOT_NAME = "franka-pick-and-place"
 ENV_NAME = "franka-pick-and-place-env"
 CARTESIAN_ENV_NAME = "franka-pick-and-place-cartesian-env"
@@ -97,6 +99,33 @@ def train(self) -> None:
         train_task(self.cfg.name, agent)
 
 
+class FrankaPickAndPlaceSkrlTask:
+    """skrl variant of :class:`FrankaPickAndPlaceTask` — same scene / reward shaping,
+    trained through the skrl backend (PPO) instead of RSL-RL. The backend is selected
+    automatically from the ``SkrlAgentCfg`` agent config."""
+
+    def __init__(self) -> None:
+        self.cfg = TaskCfg(
+            name=SKRL_TASK_ID,
+            env_name=ENV_NAME,
+            robot_name=ROBOT_NAME,
+            env=franka_pick_and_place_env_cfg(play=False),
+            play_env=franka_pick_and_place_env_cfg(play=True),
+            agent=franka_pick_and_place_skrl_cfg(),
+            trainable=True,
+        )
+
+    def play(self) -> None:
+        from genelab.rl import play_task
+
+        play_task(self.cfg.name, checkpoint=None)
+
+    def train(self) -> None:
+        from genelab.rl import train_task
+
+        train_task(self.cfg.name, self.cfg.agent)
+
+
 def register() -> None:
     if ROBOT_NAME not in ROBOTS:
         register_robot(
@@ -149,3 +178,18 @@ def register() -> None:
                 f"genelab train {CARTESIAN_TASK_ID} --num-envs 2048 --max-iterations 500",
             ],
         )
+    if SKRL_TASK_ID not in TASKS:
+        register_task(
+            SKRL_TASK_ID,
+            FrankaPickAndPlaceSkrlTask,
+            description=(
+                "skrl PPO Franka Panda pick-and-place — same task as "
+                f"{TASK_ID}, trained through the skrl backend instead of RSL-RL."
+            ),
+            cfg_type=TaskCfg,
+            examples=[
+                f"genelab play {SKRL_TASK_ID} --vis",
+                f"genelab train {SKRL_TASK_ID} --num-envs 16 --max-iterations 2",
+                f"genelab train {SKRL_TASK_ID} --num-envs 2048 --max-iterations 12000",
+            ],
+        )

pyproject.toml

@@ -25,6 +25,12 @@ rl = [
     "gymnasium>=1.0",
     "tensorboard>=2.20.0",
 ]
+# Optional second RL backend. ``genelab.rl.backends.skrl`` guards every skrl
+# import, so this extra is only needed to train/play tasks with a ``SkrlAgentCfg``.
+# Pinned ``<2`` until the skrl 2.0 model API is validated against GeneLab.
+skrl = [
+    "skrl>=1.4.0,<2",
+]
 # Live plot backends for ``genelab.recording``. PyQt plots use pyqtgraph + PyQt5;
 # Matplotlib plots use matplotlib. File-only recordings (NPZ/CSV/Video) don't need
 # any of these — they're handled by Genesis's own dependencies (numpy + PyAV).
@@ -73,6 +79,9 @@ rl = [
     "gymnasium>=1.0",
     "tensorboard>=2.20.0",
 ]
+skrl = [
+    "skrl>=1.4.0,<2",
+]
 recording = [
     "pyqtgraph>=0.13",
     "PyQt5>=5.15",
@@ -100,7 +109,7 @@ packages = ["src/genelab"]
 [tool.uv]
 # These groups install on every `uv sync` / `uv run` in this checkout, so daily
 # dev only needs to pick a torch extra (e.g. `uv sync --extra torch-cu126`).
-default-groups = ["dev", "rl", "recording", "teleop", "docs"]
+default-groups = ["dev", "rl", "skrl", "recording", "teleop", "docs"]
 conflicts = [
     [
         { extra = "torch-cpu" },

src/genelab/cli/__init__.py

@@ -542,23 +542,32 @@ def _dispatch_train(
     if agent_cfg is None:
         task.train()
         return
-    from genelab.rl import RslRlOnPolicyRunnerCfg, train_task
+    from genelab.rl import select_backend, train_task
 
-    if not isinstance(agent_cfg, RslRlOnPolicyRunnerCfg):
-        raise SystemExit(
-            f"task agent cfg has unsupported type {type(agent_cfg).__name__}; "
-            "expected RslRlOnPolicyRunnerCfg"
-        )
+    # The backend is chosen by the agent cfg type (RSL-RL, skrl, ...); an
+    # unregistered type raises a clear error here instead of deep in the runner.
+    try:
+        backend = select_backend(agent_cfg)
+    except ValueError as exc:
+        raise SystemExit(str(exc)) from exc
     task_id = getattr(task_cfg, "name", None)
     if not isinstance(task_id, str):
         raise SystemExit("task config is missing 'name'; cannot route through RL train helper")
 
     gpus_raw = runner_args.pop("gpus", None)
     gpus = int(gpus_raw) if gpus_raw is not None else 1
     num_envs_per_rank = _resolve_per_rank_num_envs(runner_args, gpus=gpus)
-    if gpus > 1 and "TORCHELASTIC_RUN_ID" not in os.environ:
-        _relaunch_under_torchrun(gpus, agent_cfg, runner_args, num_envs_per_rank, task_id=task_id)
-        return
+    if gpus > 1:
+        if backend.name != "rsl_rl":
+            raise SystemExit(
+                f"multi-GPU training (--gpus {gpus}) is only supported by the RSL-RL "
+                f"backend; this task uses the {backend.name!r} backend"
+            )
+        if "TORCHELASTIC_RUN_ID" not in os.environ:
+            _relaunch_under_torchrun(
+                gpus, agent_cfg, runner_args, num_envs_per_rank, task_id=task_id
+            )
+            return
 
     max_iter_raw = runner_args.get("max_iterations")
     seed_raw = runner_args.get("seed")

src/genelab/rl/__init__.py

@@ -1,5 +1,6 @@
-"""RL runner and VecEnv integration."""
+"""RL runner, backend abstraction, and VecEnv integration."""
 
+from genelab.rl.backends import Backend, select_backend
 from genelab.rl.config import (
     RslRlBaseRunnerCfg,
     RslRlModelCfg,
@@ -9,18 +10,24 @@
 from genelab.rl.profiler import maybe_profile, profiler_enabled
 from genelab.rl.rsl_rl_wrapper import RslRlVecEnvWrapper
 from genelab.rl.runner import AgentKind, play_task, train_task
+from genelab.rl.skrl_config import SkrlAgentCfg, SkrlExperimentCfg, SkrlModelCfg
 from genelab.rl.vecenv import VecEnvBase
 
 __all__ = [
     "AgentKind",
+    "Backend",
     "RslRlBaseRunnerCfg",
     "RslRlModelCfg",
     "RslRlOnPolicyRunnerCfg",
     "RslRlPpoAlgorithmCfg",
     "RslRlVecEnvWrapper",
+    "SkrlAgentCfg",
+    "SkrlExperimentCfg",
+    "SkrlModelCfg",
     "VecEnvBase",
     "maybe_profile",
     "play_task",
     "profiler_enabled",
+    "select_backend",
     "train_task",
 ]