feat: add RTC asynchronous execution

docs/en/configuration.md

@@ -149,6 +149,26 @@ Host. Device-only runtimes can omit the model; see
 Wireless runtimes additionally declare `inference_client.bind` and
 `inference_client.port`, and each listener on a node must use a distinct port.
 
+Asynchronous execution is opt-in through `options.async`. Use `mode: rtc` for
+Real-Time Chunking. The `overlap_steps` and `inference_delay_steps` values
+describe the modeled control window in simulation, while a physical runtime
+measures the request latency. `mode: naive` keeps the overlap baseline without
+algorithm-specific conditioning. RTC through SGLang requires its
+`/v1/actions/metadata` response to advertise the requested
+`async_inference` capability.
+
+```yaml
+runtimes:
+  libero-pi05-runtime:
+    # ... robot/simulator, model, binding, and server fields ...
+    options:
+      async:
+        mode: rtc
+        overlap_steps: 10
+        inference_delay_steps: 5 # simulation Δ; omit for a physical runtime
+        execution_horizon: 50
+```
+
 ## Providers
 
 Inference backends are registered providers. A provider declares which

docs/en/quickstart.md

@@ -121,7 +121,7 @@ uv run embodirun --config my-deployment.yaml run \
 | `--task`, `--seed` | Simulator task and episode seed. |
 | `--chunk-steps` | Actions executed from each inference chunk (capped by the binding). |
 | `--max-steps` | Maximum number of inference/action chunks (default 1). |
-| `--control-hz` | Action playback rate (default 5). |
+| `--control-hz` | Action playback rate (default 5); set to `0` to disable software pacing. |
 | `--request-timeout` | Timeout for one inference request (default 60 s). |
 
 A π0.5 response has no task-complete signal. The chunk bound limits the run;

docs/zh/configuration.md

@@ -146,6 +146,24 @@ runtimes:
 无线运行时还需声明 `inference_client.bind` 和
 `inference_client.port`，并且节点上的每个监听器必须使用不同的端口。
 
+异步执行通过 `options.async` 显式开启。`mode: rtc` 使用 Real-Time Chunking。
+仿真器用 `overlap_steps` 和 `inference_delay_steps` 描述模拟的控制窗口，真机运行时会测量实际请求延迟。
+`mode: naive` 可用作不携带算法条件的异步基线。通过 SGLang
+使用 RTC 时，其 `/v1/actions/metadata` 必须声明对应的
+`async_inference` 能力。
+
+```yaml
+runtimes:
+  libero-pi05-runtime:
+    # ... simulator/robot、model、binding 和 server 字段 ...
+    options:
+      async:
+        mode: rtc
+        overlap_steps: 10
+        inference_delay_steps: 5 # 仿真 Δ；真机可以省略
+        execution_horizon: 50
+```
+
 ## 扩展推理后端：provider {#provider}
 
 每个推理后端通过 provider 注册，声明支持的传输协议、启动命令、依赖组和模型参数。

docs/zh/quickstart.md

@@ -111,7 +111,7 @@ uv run embodirun --config my-deployment.yaml run \
 | `--task`、`--seed` | 仿真器任务与 episode 种子。 |
 | `--chunk-steps` | 每个动作块中实际执行的步数，不能超过绑定上限。 |
 | `--max-steps` | 推理与动作执行的最大轮数（默认 1）。 |
-| `--control-hz` | 动作回放频率（默认 5）。 |
+| `--control-hz` | 动作回放频率（默认 5）；设为 `0` 可关闭软件节拍。 |
 | `--request-timeout` | 单次推理请求的超时时间（默认 60 s）。 |
 
 π0.5 每次返回一个动作块（chunk），不返回任务完成信号。运行在达到设定的动作块数量后结束，

integrations/sglang_pi05/README.md

@@ -4,3 +4,10 @@ This optional package owns the LeRobot Pi05 checkpoint compatibility pipeline.
 Install it only in the managed SGLang inference environment. The managed entry
 point is `embodirun-sglang-pi05-serve`. EmbodiRun's own
 `embodirun-sglang-serve` launcher stays a thin wrapper in the root package.
+
+The Pi05 pipeline exposes the shared `embodiinfer.async.v1` RTC metadata
+contract. It keeps a per-session model-space action chunk and applies the
+requested hard-prefix or soft-VJP denoising correction. RTC is currently
+supported for SGLang's
+monolithic worker (the normal single-4090 deployment); a split prefix/action
+worker rejects an RTC request instead of silently running it synchronously.