KraHsu · May 29, 2026 · May 29, 2026
diff --git a/docs/cli/list-info.en.md b/docs/cli/list-info.en.md
@@ -11,37 +11,51 @@ genelab list envs
 genelab list tasks
 ```
 
-If an extension is not installed, import it explicitly:
+If an extension is not installed as an entry point, import it explicitly:
 
 ```bash
 PYTHONPATH=examples/inverted_pendulum/src \
   genelab --import genelab_inverted_pendulum.tasks list tasks
 ```
 
-## Inspect a task
+To see only the bundled asset zoo (skip user-installed entry-point extensions):
 
 ```bash
-genelab info GeneLab-Inverted-Pendulum-v0
+genelab --no-entry-points list robots
 ```
 
-The task view includes:
+## Inspect any registered name
+
+```bash
+genelab info GeneLab-Inverted-Pendulum-v0   # task
+genelab info CartPole-Env-v0                # env
+genelab info anymal_c                       # robot
+```
+
+`info` does not require you to know which registry the name belongs to. Task views additionally
+include:
 
 | Section | Meaning |
 |---|---|
 | Metadata | Registered name, description, cfg type, examples. |
 | Task config | `TaskCfg` summary: env, robot, trainable flag, agent. |
 | Overridable paths | Dotted paths accepted by `play` and `train`. |
 
+If the inspected task needs an asset that isn't cached yet, `info` fetches it inline behind the
+same progress bar used by `play` / `train` / `asset download`.
+
 ## Use copied override paths
 
 ```bash
 genelab play GeneLab-Inverted-Pendulum-v0 \
   --env.rewards_cfg.pole_upright.weight 4.0
 ```
 
-When `play_env` exists, play-mode shortcut flags target `play_env`; explicit paths remain explicit.
+When `play_env` exists, play-mode shortcut flags (`--vis` / `--gpu` / `--dt` / `--steps`) target
+`play_env`; explicit dotted paths still write the exact subtree you spelled.
 
 ## See also
 
 - [Configuration Reference](../reference/configuration.md)
 - [CLI Reference](../reference/cli.md)
+- [Asset zoo](../concepts/asset_zoo.md)
diff --git a/docs/cli/list-info.zh.md b/docs/cli/list-info.zh.md
@@ -10,37 +10,50 @@ genelab list envs
 genelab list tasks
 ```
 
-如果扩展尚未安装，显式导入：
+如果扩展尚未通过 entry point 安装，显式导入：
 
 ```bash
 PYTHONPATH=examples/inverted_pendulum/src \
   genelab --import genelab_inverted_pendulum.tasks list tasks
 ```
 
-## 查看任务
+要只看随包 asset zoo（不加载用户已安装的扩展），加 `--no-entry-points`：
 
 ```bash
-genelab info GeneLab-Inverted-Pendulum-v0
+genelab --no-entry-points list robots
 ```
 
-任务视图包括：
+## 查看任意已注册名字
+
+```bash
+genelab info GeneLab-Inverted-Pendulum-v0   # task
+genelab info CartPole-Env-v0                # env
+genelab info anymal_c                       # robot
+```
+
+`info` 不要求事先知道是哪一类。Task 视图额外包含：
 
 | 区域 | 含义 |
 |---|---|
 | 元信息 | 注册名、描述、cfg 类型、示例命令。 |
 | 任务配置 | `TaskCfg` 摘要：env、robot、trainable、agent。 |
 | 可覆盖路径 | `play` 和 `train` 接受的 dotted path。 |
 
+若该 task 需要的 asset 尚未缓存，`info` 会即时下载并显示进度条
+（和 `play` / `train` / `asset download` 走的是同一条 progress UI）。
+
 ## 使用复制出的 override 路径
 
 ```bash
 genelab play GeneLab-Inverted-Pendulum-v0 \
   --env.rewards_cfg.pole_upright.weight 4.0
 ```
 
-当 task 有 `play_env` 时，play 模式短标志会作用到 `play_env`；显式路径仍保持显式语义。
+当 task 提供 `play_env` 时，play 模式下的简写 flag（`--vis` / `--gpu` / `--dt` /
+`--steps`）会作用到 `play_env`；显式的 dotted 路径仍按字面写的子树生效。
 
 ## 另见
 
 - [配置参考](../reference/configuration.md)
 - [CLI 参考](../reference/cli.md)
+- [Asset zoo](../concepts/asset_zoo.md)
diff --git a/docs/cli/overview.en.md b/docs/cli/overview.en.md
@@ -12,10 +12,21 @@ genelab [global options] <command> [arguments]
 
 | Area | Commands |
 |---|---|
-| Registry discovery | `list robots`, `list envs`, `list tasks`, `info NAME` |
-| Runtime | `play TASK`, `train TASK` |
+| Registry | `list robots\|envs\|tasks`, `info NAME`, `asset list\|info\|download\|purge` |
+| Runtime | `play TASK`, `train TASK`, `eval TASK CHECKPOINT`, `export TASK CHECKPOINT`, `benchmark --suite ...` |
+| Project | `project new NAME` |
 | Utilities | `cache`, `prof open` |
-| Project scaffolding | `project new NAME` |
+
+Running `genelab` with no subcommand prints a landing page with quickstart commands and a count of
+registered robots / envs / tasks.
+
+## Global options
+
+| Option | Meaning |
+|---|---|
+| `--version` | Print the version and exit. |
+| `--import MODULE` | Import an extension module before dispatch; repeatable. |
+| `--no-entry-points` | Skip extensions from the `genelab.extensions` entry-point group. |
 
 ## Extension loading order
 
@@ -30,7 +41,7 @@ Use entry points for daily work and `--import` for local experiments.
 ## Overrides
 
 Runtime commands accept unknown `--a.b.c VALUE` options after the task id. The CLI forwards them to
-`apply_overrides`.
+`apply_overrides`:
 
 ```bash
 genelab play TASK_ID --env.simulation.dt 0.005
@@ -48,3 +59,4 @@ When stdin is a TTY, the CLI can prompt for a missing task id, unknown registry
 - [CLI Reference](../reference/cli.md)
 - [Discovery: list and info](list-info.md)
 - [Play and Train](play-train.md)
+- [Eval and export](../concepts/eval-and-export.md)
diff --git a/docs/cli/overview.zh.md b/docs/cli/overview.zh.md
@@ -1,47 +1,61 @@
 # CLI 总览
 
-`genelab` CLI 是注册表和任务配置上的薄调度层。它不拥有任务逻辑；它负责发现扩展、解析已注册对象、应用 override，然后调用 task 或 runner。
+`genelab` CLI 是注册表（registry）与任务配置之上的薄分发层。它不持有 task 逻辑——
+发现扩展、解析已注册对象、应用 override，然后调用 task 或 runner。
 
 ## 命令模型
 
 ```bash
 genelab [全局选项] <命令> [参数]
 ```
 
-| 区域 | 命令 |
+| 分组 | 命令 |
 |---|---|
-| 注册表发现 | `list robots`、`list envs`、`list tasks`、`info NAME` |
-| 运行时 | `play TASK`、`train TASK` |
-| 工具 | `cache`、`prof open` |
-| 项目骨架 | `project new NAME` |
+| Registry | `list robots\|envs\|tasks`、`info NAME`、`asset list\|info\|download\|purge` |
+| Runtime | `play TASK`、`train TASK`、`eval TASK CHECKPOINT`、`export TASK CHECKPOINT`、`benchmark --suite ...` |
+| Project | `project new NAME` |
+| Utilities | `cache`、`prof open` |
+
+不带子命令运行 `genelab` 会打印 landing 页：包含 quickstart 命令以及已注册的
+robots / envs / tasks 数量。
+
+## 全局选项
+
+| 选项 | 含义 |
+|---|---|
+| `--version` | 打印版本并退出。 |
+| `--import MODULE` | 分发前导入扩展模块；可重复。 |
+| `--no-entry-points` | 跳过 `genelab.extensions` entry point 组的扩展。 |
 
 ## 扩展加载顺序
 
-任何需要注册表数据的命令都会按顺序加载：
+每个需要 registry 数据的命令按以下顺序加载扩展：
 
-1. 通过 `load_bundled_asset_zoo()` 加载随包 asset zoo robot。
-2. 已安装的 `genelab.extensions` entry point，除非设置 `--no-entry-points`。
-3. 重复传入的显式 `--import MODULE`。
+1. 通过 `load_bundled_asset_zoo()` 装入随包的 asset zoo 机器人。
+2. 已安装的 `genelab.extensions` entry points（除非传 `--no-entry-points`）。
+3. 多次显式的 `--import MODULE`。
 
-日常工作用 entry point，本地实验用 `--import`。
+日常工作用 entry points，临时实验用 `--import`。
 
 ## Overrides
 
-运行时命令在 task id 后接受未知的 `--a.b.c VALUE` 选项。CLI 会把它们转发给
-`apply_overrides`。
+Runtime 命令在 task id 后接受未知的 `--a.b.c VALUE` 选项，由 CLI 转发到
+`apply_overrides`：
 
 ```bash
 genelab play TASK_ID --env.simulation.dt 0.005
 ```
 
-用 `genelab info TASK_ID` 查看有效路径。
+可用路径见 `genelab info TASK_ID`。
 
 ## 交互模式
 
-stdin 是 TTY 时，CLI 可以为缺失 task id、未知注册表名字、非法 `--agent`、未知 override 路径弹出选择器。CI 和脚本中会直接抛出相同错误。
+当 stdin 是 TTY 时，CLI 可以为缺失的 task id、未知的注册名、非法 `--agent` 或
+未知 override 路径弹出交互选择器；CI 与脚本中则直接抛出同样的错误。
 
 ## 另见
 
 - [CLI 参考](../reference/cli.md)
 - [发现：list 与 info](list-info.md)
 - [play 与 train](play-train.md)
+- [Eval 与 export](../concepts/eval-and-export.md)
diff --git a/docs/cli/play-train.en.md b/docs/cli/play-train.en.md
@@ -1,7 +1,8 @@
-# Play and Train
+# Runtime: play and train
 
 `play` runs a registered task. `train` runs a registered task through a supported runner when the
-task provides an agent config.
+task provides an agent config. This page also covers the post-training runtime subcommands —
+`eval`, `export`, and `benchmark` — which take the same checkpoint produced by `train`.
 
 ## Play
 
@@ -23,8 +24,8 @@ The policy options (`--agent`, `--checkpoint`, `--num-envs`, `--prof*`) apply on
 tasks — those whose play env config is a `ManagerBasedRlEnvCfg`. Non-RL **scene-playback
 demos** (e.g. `GeneLab-Rubiks-Play-v0`, `GeneLab-Wuji-Hand-Playback-v0`), whose config
 subclasses the base `ManagerBasedEnvCfg`, run their own built-in playback; passing those
-options prints a warning and they are ignored. `--steps` / `--vis` / `--gpu` and dotted
-config overrides still apply to both.
+options prints a warning and they are ignored. `--steps` / `--vis` / `--headless` / `--gpu` /
+`--dt` and dotted config overrides still apply to both.
 
 Checkpoint replay:
 
@@ -44,6 +45,19 @@ genelab play TASK_ID \
       --checkpoint <ckpt> --headless
     ```
 
+## Shortcut flags
+
+Both `play` and `train` rewrite the following shortcuts into `env.simulation.*` overrides:
+
+| Shortcut | Override |
+|---|---|
+| `-v`, `--vis` | `env.simulation.vis=true` |
+| `--headless` | `env.simulation.vis=false` (mutually exclusive with `--vis`) |
+| `--gpu` | `env.simulation.gpu=true` |
+| `--steps N` | play: `env.simulation.steps=N`; train: alias for `--max_iterations N` |
+| `--dt SECONDS` | `env.simulation.dt=SECONDS` |
+| `--a.b.c VALUE` | Any dotted cfg path |
+
 ## Train
 
 ```bash
@@ -57,7 +71,38 @@ 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.
+`--num_envs_per_gpu` for per-rank semantics (mutually exclusive with `--num_envs`).
+Multi-GPU is RSL-RL only; the first entry automatically relaunches under `torchrun`.
+
+### In-training eval
+
+Pass `--eval_every K` to run a deterministic rollout every K iterations. On improvement, the
+runner writes `best_model.<ext>` into `--log_dir`:
+
+```bash
+genelab train TASK_ID --eval_every 50 --eval_episodes 20
+```
+
+| Option | Meaning (default) |
+|---|---|
+| `--eval_every K` | Evaluate every K iterations. |
+| `--eval_episodes N` | Episodes per evaluation (10). |
+| `--eval_num_envs N` | Parallel envs during eval (matches training). |
+| `--eval_seed N` | RNG seed for the eval rollout (0). |
+
+### Multi-seed train
+
+`--seeds 1,2,3` fans out the current `train` invocation into one independent subprocess per seed:
+
+```bash
+genelab train TASK_ID --seeds 1,2,3,4 --parallel 2 --num_envs 4096
+```
+
+- `--parallel N` caps concurrency (default 1 — sequential).
+- Each child is invoked with `--seed S` and `--log_dir <parent>/seed_<S>`.
+- Without an explicit `--log_dir`, the parent is
+  `logs/multi-seed/<task_id>/<YYYY-MM-DD_HH-MM-SS>/`.
+- If any seed fails, the command exits non-zero.
 
 ## RL backends
 
@@ -95,6 +140,47 @@ GENELAB_SB3_DEMO_PATH=/tmp/franka_pp_demos.npz \
   --gpu --num-envs 32 --max-iterations 2000000
 ```
 
+## Post-training subcommands
+
+`eval`, `export`, and `benchmark` all take a registered task plus a checkpoint and reuse the
+task's play env config — see [Eval and export](../concepts/eval-and-export.md) for the
+conceptual picture and script-level details.
+
+### Eval
+
+Deterministic rollout that writes `eval.json` (`return_mean`, `length_mean`, and
+`success_rate` if the task publishes `extras['is_success']`):
+
+```bash
+genelab eval TASK_ID logs/.../model_300.pt \
+  --num-envs 64 --episodes 100 --out eval.json
+```
+
+`--deterministic` / `--stochastic` toggles the policy mode; `--max-steps` caps the rollout.
+
+### Export
+
+Export the policy as TorchScript or ONNX (per-term scale/clip baked into the model):
+
+```bash
+genelab export TASK_ID logs/.../model_300.pt --format onnx --out policy.onnx
+```
+
+A sibling `<OUTPUT>.metadata.json` records the observation schema.
+
+### Benchmark
+
+Batch eval driven by a JSON suite, aggregated into one report:
+
+```bash
+genelab benchmark --suite suite.json --out report.json
+genelab benchmark --suite suite.json --reference baseline.json --tolerance 0.1
+```
+
+`suite.json` is `[{"task": ..., "checkpoint": ..., "episodes": ..., "seed": ..., "num_envs": ...}, ...]`.
+With `--reference`, the command compares `return_mean` against the baseline and exits non-zero
+when any task drops more than `--tolerance` — usable directly as a CI regression gate.
+
 ## Config overrides
 
 Any unknown option after the task id is treated as a dotted config override:
@@ -110,3 +196,4 @@ genelab play TASK_ID \
 - [Run RL Experiments](../best-practices/rl-experiments.md)
 - [CLI Reference](../reference/cli.md)
 - [RL runner](../concepts/rl-runner.md)
+- [Eval and export](../concepts/eval-and-export.md)