feat(rollout): interactive v1

docs/source/inference.mdx

@@ -243,22 +243,23 @@ See the [Real-Time Chunking](./rtc) guide for details on tuning RTC parameters.
 
 ## Common Flags
 
-| Flag                              | Description                                                       | Default |
+| Flag                              | Description                                                                                                     | Default |
-| --------------------------------- | ----------------------------------------------------------------- | ------- |
+| --------------------------------- | --------------------------------------------------------------------------------------------------------------- | ------- |
-| `--policy.path`                   | **Required.** HF Hub model ID or local checkpoint path            | --      |
+| `--policy.path`                   | **Required.** HF Hub model ID or local checkpoint path                                                          | --      |
-| `--robot.type`                    | **Required.** Robot type (e.g. `so100_follower`, `koch_follower`) | --      |
+| `--robot.type`                    | **Required.** Robot type (e.g. `so100_follower`, `koch_follower`)                                               | --      |
-| `--robot.port`                    | Serial port for the robot                                         | --      |
+| `--robot.port`                    | Serial port for the robot                                                                                       | --      |
-| `--robot.cameras`                 | Camera configuration (JSON dict)                                  | --      |
+| `--robot.cameras`                 | Camera configuration (JSON dict)                                                                                | --      |
-| `--fps`                           | Control loop frequency                                            | 30      |
+| `--fps`                           | Control loop frequency                                                                                          | 30      |
-| `--duration`                      | Run time in seconds (0 = infinite)                                | 0       |
+| `--duration`                      | Run time in seconds (0 = infinite)                                                                              | 0       |
-| `--device`                        | Torch device (`cpu`, `cuda`, `mps`)                               | auto    |
+| `--device`                        | Torch device (`cpu`, `cuda`, `mps`)                                                                             | auto    |
-| `--task`                          | Task description (used when no dataset is provided)               | --      |
+| `--task`                          | Task description (used when no dataset is provided)                                                             | --      |
-| `--display_data`                  | Stream telemetry to Rerun visualization                           | false   |
+| `--display_data`                  | Stream telemetry to Rerun visualization                                                                         | false   |
-| `--display_ip` / `--display_port` | Remote Rerun server address                                       | --      |
+| `--display_ip` / `--display_port` | Remote Rerun server address                                                                                     | --      |
-| `--interpolation_multiplier`      | Action interpolation factor                                       | 1       |
+| `--interpolation_multiplier`      | Action interpolation factor                                                                                     | 1       |
-| `--use_torch_compile`             | Enable `torch.compile` for inference                              | false   |
+| `--interactive`                   | Chat-style stdin session (`/start`, `/reset`, `/stop`); the robot stays idle until `/start`. Base strategy only | false   |
-| `--resume`                        | Resume a previous recording session                               | false   |
+| `--use_torch_compile`             | Enable `torch.compile` for inference                                                                            | false   |
-| `--play_sounds`                   | Vocal synthesis for events                                        | true    |
+| `--resume`                        | Resume a previous recording session                                                                             | false   |
+| `--play_sounds`                   | Vocal synthesis for events                                                                                      | true    |
 
 ---
 

src/lerobot/rollout/__init__.py

@@ -47,6 +47,13 @@ from .inference import (
     SyncInferenceEngine,
     create_inference_engine,
 )
+from .interactive import (
+    InteractiveCommand,
+    InteractiveSession,
+    LinkedEvent,
+    StdinCommandListener,
+    parse_command,
+)
 from .strategies import (
     BaseStrategy,
     DAggerStrategy,
@@ -65,13 +72,16 @@ __all__ = [
     "DAggerStrategy",
     "DAggerStrategyConfig",
     "DatasetContext",
+    "EpisodicStrategy",
+    "EpisodicStrategyConfig",
     "HardwareContext",
     "HighlightStrategy",
     "HighlightStrategyConfig",
-    "EpisodicStrategy",
-    "EpisodicStrategyConfig",
     "InferenceEngine",
     "InferenceEngineConfig",
+    "InteractiveCommand",
+    "InteractiveSession",
+    "LinkedEvent",
     "PolicyContext",
     "ProcessorContext",
     "RTCInferenceConfig",
@@ -83,9 +93,11 @@ __all__ = [
     "RuntimeContext",
     "SentryStrategy",
     "SentryStrategyConfig",
+    "StdinCommandListener",
     "SyncInferenceConfig",
     "SyncInferenceEngine",
     "build_rollout_context",
     "create_inference_engine",
     "create_strategy",
+    "parse_command",
 ]

src/lerobot/rollout/configs.py

@@ -239,6 +239,11 @@ class RolloutConfig:
     # Runtime
     fps: float = 30.0
     duration: float = 0.0  # 0 = infinite (24/7 mode)
+    # Interactive session: control the rollout from stdin with chat-style
+    # commands (/start, /reset, /stop) while hardware and policy stay warm.
+    # The robot does not move until /start is received.  Currently limited to
+    # --strategy.type=base.
+    interactive: bool = False
     interpolation_multiplier: int = 1
     device: str | None = None
     task: str = ""
@@ -294,6 +299,17 @@ class RolloutConfig:
                 "Base strategy does not record data. Use sentry, highlight, or dagger for recording."
             )
 
+        # Interactive mode drives strategy.run() in restartable segments and reads
+        # commands from stdin.  Recording strategies are excluded for now: their
+        # run() loops finalize the dataset on exit (so they cannot be restarted)
+        # and their keyboard listeners read the same terminal as the command
+        # prompt.
+        if self.interactive and not isinstance(self.strategy, BaseStrategyConfig):
+            raise ValueError(
+                f"--interactive=true currently supports only --strategy.type=base "
+                f"(got '{self.strategy.type}')."
+            )
+
         # Sentry MUST use streaming encoding to avoid disk I/O blocking the control loop
         if (
             isinstance(self.strategy, SentryStrategyConfig)

src/lerobot/rollout/inference/rtc.py

@@ -140,6 +140,9 @@ class RTCInferenceEngine(InferenceEngine):
         self._action_queue: ActionQueue | None = None
         self._obs_holder: dict[str, Any] = {}
         self._obs_lock = Lock()
+        # Bumped by reset() (under _obs_lock) so chunks whose inference started
+        # before a reset are discarded instead of merged into the fresh queue.
+        self._reset_epoch = 0
         self._policy_active = Event()
         self._compile_warmup_done = Event()
         self._shutdown_event = Event()
@@ -235,13 +238,26 @@ class RTCInferenceEngine(InferenceEngine):
         self._policy_active.set()
 
     def reset(self) -> None:
-        """Reset the policy, processors, and action queue."""
+        """Reset the policy, processors, and action queue.
+
+        Call while the engine is paused (both DAgger transitions and the
+        interactive session do): the RTC thread may still be finishing an
+        inference started before the pause, so ``reset`` also drops the last
+        published observation — it can be arbitrarily stale by the time the
+        engine resumes (e.g. the robot was returned to its initial position
+        in the meantime), and a chunk computed from it would jerk the robot
+        toward the old pose — and bumps the reset epoch so any in-flight
+        chunk is discarded instead of merged into the cleared queue.
+        """
         logger.info("Resetting RTC inference state (policy + processors + queue)")
         self._policy.reset()
         self._preprocessor.reset()
         self._postprocessor.reset()
         if self._action_queue is not None:
             self._action_queue.clear()
+        with self._obs_lock:
+            self._obs_holder["obs"] = None
+            self._reset_epoch += 1
 
     # ------------------------------------------------------------------
     # Action production (called from main thread)
@@ -281,6 +297,7 @@ class RTCInferenceEngine(InferenceEngine):
                 queue = self._action_queue
                 with self._obs_lock:
                     obs = self._obs_holder.get("obs")
+                    epoch_before = self._reset_epoch
                 if queue is None or obs is None:
                     time.sleep(_RTC_IDLE_SLEEP_S)
                     continue
@@ -339,7 +356,12 @@ class RTCInferenceEngine(InferenceEngine):
                         else:
                             latency_tracker.add(new_latency)
 
-                        queue.merge(original, processed, new_delay, idx_before)
+                        with self._obs_lock:
+                            epoch_unchanged = epoch_before == self._reset_epoch
+                        if epoch_unchanged:
+                            queue.merge(original, processed, new_delay, idx_before)
+                        else:
+                            logger.info("Discarding action chunk computed before an engine reset")
 
                         if (
                             is_warmup

src/lerobot/rollout/interactive.py

@@ -0,0 +1,474 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Interactive rollout session: chat-style stdin commands for ``lerobot-rollout``.
+
+Enabled with ``--interactive=true``, this module lets the operator control a
+rollout from the terminal while hardware and policy stay connected and warm:
+
+    /start   start (or restart) the policy control loop
+    /reset   stop movement and return the robot to its initial position
+    /stop    end the session and run the normal shutdown routines
+    /help    show the available commands
+
+Threading model (mirrors the DAgger events pattern): a daemon
+:class:`StdinCommandListener` thread reads lines and only ever sets
+thread-safe flags — it never touches hardware or the inference engine.  The
+:class:`InteractiveSession` driver runs on the main thread and executes
+``strategy.run(ctx)`` in *segments*: each ``/start`` begins a segment, and
+``/reset`` / ``/stop`` end it by setting the session's :class:`LinkedEvent`,
+which every strategy control loop already polls as
+``ctx.runtime.shutdown_event``.  Real shutdown signals (SIGINT/SIGTERM)
+propagate through the linked event's parent, so Ctrl-C behaves exactly as in
+non-interactive runs.
+
+The command table is intentionally a name → handler mapping so future
+commands (``/subtask``, ``/ask`` — see the language-runtime work in
+PR #4183/#4234) can be registered without restructuring the parser or the
+session loop.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import select
+import sys
+import time
+from collections.abc import Callable
+from dataclasses import dataclass
+from threading import Event, Thread
+from typing import IO, TYPE_CHECKING
+
+from lerobot.utils.utils import log_say
+
+if TYPE_CHECKING:
+    from .context import RolloutContext
+    from .strategies import RolloutStrategy
+
+logger = logging.getLogger(__name__)
+
+_BANNER_RULE = "─" * 60
+
+
+class LinkedEvent(Event):
+    """A ``threading.Event`` whose ``is_set`` also reflects a parent event.
+
+    ``set``/``clear`` act only on the local flag, so the interactive session
+    can raise and clear its own segment-stop requests without masking (or
+    accidentally re-arming) the process-wide shutdown event carried by
+    ``parent``.  Every rollout strategy control loop polls
+    ``ctx.runtime.shutdown_event.is_set()``, so installing a ``LinkedEvent``
+    there makes the loops react both to session commands and to real
+    shutdown signals.
+    """
+
+    _WAIT_SLICE_S = 0.05
+
+    def __init__(self, parent: Event) -> None:
+        super().__init__()
+        self.parent = parent
+
+    def is_set(self) -> bool:
+        return super().is_set() or self.parent.is_set()
+
+    def wait(self, timeout: float | None = None) -> bool:
+        """Wait for either the local or the parent flag.
+
+        The base ``Event.wait`` only watches the local flag, so poll in short
+        slices to also observe the parent.  Strategy loops only call
+        ``is_set()``; this coarse wait exists for API completeness.
+        """
+        deadline = None if timeout is None else time.perf_counter() + timeout
+        while not self.is_set():
+            remaining = None if deadline is None else deadline - time.perf_counter()
+            if remaining is not None and remaining <= 0:
+                return False
+            wait_slice = self._WAIT_SLICE_S if remaining is None else min(self._WAIT_SLICE_S, remaining)
+            super().wait(wait_slice)
+        return True
+
+
+@dataclass(frozen=True)
+class InteractiveCommand:
+    """A parsed ``/name args`` line from the interactive prompt."""
+
+    name: str
+    args: str = ""
+
+
+def parse_command(line: str) -> InteractiveCommand | None:
+    """Parse an input line into an :class:`InteractiveCommand`.
+
+    Commands are ``/name`` optionally followed by free-text arguments
+    (unused by the built-in commands, but the grammar already supports
+    future ones like ``/subtask grab the red cube``).  Returns ``None`` for
+    lines that are not commands (no leading ``/`` or a bare ``/``).
+    """
+    line = line.strip()
+    if not line.startswith("/"):
+        return None
+    head, *rest = line.split(maxsplit=1)
+    name = head[1:].lower()
+    if not name:
+        return None
+    return InteractiveCommand(name=name, args=rest[0].strip() if rest else "")
+
+
+class StdinCommandListener:
+    """Daemon thread that reads input lines and forwards them to a callback.
+
+    On POSIX the reader polls the stream with ``select`` so ``stop()`` can
+    end the thread promptly; elsewhere (or for file-like objects without a
+    file descriptor) it falls back to a blocking ``readline`` daemon thread
+    that dies with the process.  Blank lines are skipped; end-of-file and
+    unexpected read errors trigger ``on_eof`` (an interactive Ctrl-D or an
+    exhausted piped script both mean "no more commands" — the session must
+    not keep the robot running with no way to command it).
+
+    Unlike :class:`lerobot.utils.keyboard_input.TerminalKeyListener`, this
+    reader leaves the terminal in canonical (line-buffered, echoing) mode —
+    the operator is typing chat-style commands, not pressing hotkeys.
+    """
+
+    def __init__(
+        self,
+        on_line: Callable[[str], None],
+        on_eof: Callable[[], None] | None = None,
+        stream: IO[str] | None = None,
+        poll_interval_s: float = 0.2,
+    ) -> None:
+        self._on_line = on_line
+        self._on_eof = on_eof
+        self._stream = stream if stream is not None else sys.stdin
+        self._poll_interval_s = poll_interval_s
+        self._running = False
+        self._thread: Thread | None = None
+        self._use_select = False
+        if os.name == "posix":
+            try:
+                self._stream.fileno()
+                self._use_select = True
+            except (OSError, ValueError, AttributeError):
+                pass
+
+    def start(self) -> None:
+        """Start the reader thread (idempotent)."""
+        if self._thread is not None:
+            return
+        self._running = True
+        self._thread = Thread(target=self._run, daemon=True, name="InteractiveStdin")
+        self._thread.start()
+        if not self._use_select:
+            logger.info("stdin listener running in blocking mode (select unavailable for this stream)")
+
+    def stop(self) -> None:
+        """Stop the reader thread.
+
+        Blocking-mode threads may be stuck inside ``readline`` and cannot be
+        joined; they are daemons and die with the process. Late lines are
+        ignored via the ``_running`` flag either way.
+        """
+        self._running = False
+        thread = self._thread
+        self._thread = None
+        if thread is not None and thread.is_alive() and self._use_select:
+            thread.join(timeout=1.0)
+
+    def _run(self) -> None:
+        if self._use_select:
+            self._run_select()
+        else:
+            self._run_blocking()
+
+    def _run_select(self) -> None:
+        """Poll the file descriptor and split lines from raw bytes.
+
+        Reading raw bytes (instead of ``stream.readline()``) matters: a
+        buffered file object can slurp several lines off the descriptor at
+        once, after which ``select`` reports the drained fd as not-ready and
+        the buffered lines would never be delivered — breaking pasted or
+        piped command sequences.
+        """
+        fd = self._stream.fileno()
+        buffer = b""
+        while self._running:
+            try:
+                ready, _, _ = select.select([fd], [], [], self._poll_interval_s)
+            except (OSError, ValueError):  # stream closed underneath us
+                self._emit_read_error()
+                return
+            if not ready:
+                continue
+            try:
+                chunk = os.read(fd, 4096)
+            except OSError:
+                self._emit_read_error()
+                return
+            if not self._running:
+                return
+            if chunk == b"":  # EOF: Ctrl-D or the piped input ended
+                self._emit_line(buffer)  # a final command without trailing newline still counts
+                self._emit_eof()
+                return
+            buffer += chunk
+            while b"\n" in buffer:
+                raw, buffer = buffer.split(b"\n", 1)
+                self._emit_line(raw)
+
+    def _run_blocking(self) -> None:
+        while self._running:
+            try:
+                line = self._stream.readline()
+            except (OSError, ValueError):
+                self._emit_read_error()
+                return
+            if not self._running:
+                return
+            if line == "":  # EOF
+                self._emit_eof()
+                return
+            self._emit_line(line.encode() if isinstance(line, str) else line)
+
+    def _emit_line(self, raw: bytes) -> None:
+        line = raw.decode(errors="replace").strip()
+        if not line:
+            return
+        try:
+            self._on_line(line)
+        except Exception:  # never let a handler error kill the reader thread
+            logger.exception("Error while handling interactive input %r", line)
+
+    def _emit_eof(self) -> None:
+        logger.info("Interactive input stream closed (EOF)")
+        if self._on_eof is not None:
+            try:
+                self._on_eof()
+            except Exception:
+                logger.exception("Error while handling interactive input EOF")
+
+    def _emit_read_error(self) -> None:
+        """Treat an unexpected read failure like EOF so the session shuts down.
+
+        A dead command channel must not leave the robot running with no way
+        to stop it.  Deliberate ``stop()`` calls clear ``_running`` first and
+        do not reach this path.
+        """
+        if self._running:
+            logger.warning("Interactive input stream failed — treating as EOF")
+            self._emit_eof()
+
+
+class InteractiveSession:
+    """Drive a rollout strategy from chat-style stdin commands.
+
+    The session owns the outer lifecycle: after ``strategy.setup(ctx)`` the
+    robot stays idle until ``/start``.  Each run *segment* executes
+    ``strategy.run(ctx)`` on the calling (main) thread until the operator
+    interrupts it or the strategy returns on its own (e.g. ``--duration``
+    elapsed).  ``/reset`` pauses the inference engine and returns the robot
+    to its initial position while hardware and policy stay warm; ``/stop``
+    ends the session so the caller can run ``strategy.teardown(ctx)`` — the
+    same shutdown routine as non-interactive rollouts.
+
+    Requires ``ctx.runtime.shutdown_event`` to be a :class:`LinkedEvent`
+    (installed by ``lerobot-rollout`` when ``--interactive=true``): the
+    session sets the local flag to end a segment, and process signals still
+    propagate through the parent.
+
+    Commands are last-write-wins: ``/reset`` and ``/stop`` cancel a pending
+    ``/start`` so the robot never starts moving after the operator's final
+    command asked it not to.  End-of-file on the command stream stops the
+    session (a closed stdin means there is no way left to command the
+    robot), so piped scripts must keep stdin open for the intended session
+    duration, e.g. ``(printf '/start\\n'; sleep 60; printf '/stop\\n') |
+    lerobot-rollout ... --interactive=true``.
+    """
+
+    _POLL_INTERVAL_S = 0.2
+
+    def __init__(
+        self,
+        strategy: RolloutStrategy,
+        ctx: RolloutContext,
+        input_stream: IO[str] | None = None,
+    ) -> None:
+        stop_event = ctx.runtime.shutdown_event
+        if not isinstance(stop_event, LinkedEvent):
+            raise TypeError(
+                "InteractiveSession requires ctx.runtime.shutdown_event to be a LinkedEvent so "
+                "/reset can end a run segment without triggering process shutdown. Build the "
+                "rollout context with build_rollout_context(cfg, LinkedEvent(shutdown_event))."
+            )
+        self._strategy = strategy
+        self._ctx = ctx
+        self._segment_stop = stop_event
+        self._global_shutdown = stop_event.parent
+        self._listener = StdinCommandListener(self._handle_line, on_eof=self._handle_eof, stream=input_stream)
+
+        # Written by the listener thread, consumed by the main loop.
+        self._start_requested = Event()
+        self._reset_requested = Event()
+        self._stop_requested = Event()
+        self._wake = Event()
+        self._running = Event()
+
+        # name -> (handler, help line); /help and the banner render from this
+        # table, so future commands (/subtask, /ask) stay documented for free.
+        self._commands: dict[str, tuple[Callable[[InteractiveCommand], None], str]] = {
+            "start": (self._cmd_start, "start (or restart) the policy control loop"),
+            "reset": (self._cmd_reset, "stop movement and return the robot to its initial position"),
+            "stop": (self._cmd_stop, "end the session and shut down"),
+            "help": (self._cmd_help, "show this help"),
+        }
+
+    # ------------------------------------------------------------------
+    # Main-thread session loop
+    # ------------------------------------------------------------------
+
+    def run(self) -> None:
+        """Run the session until ``/stop``, EOF, engine failure, or a shutdown signal."""
+        play_sounds = self._ctx.runtime.cfg.play_sounds
+        self._print(self._render_banner())
+        self._listener.start()
+        try:
+            while not self._global_shutdown.is_set():
+                if self._ctx.policy.inference.failed:
+                    self._print("Inference engine failed — shutting down. See the log for the error.")
+                    break
+                if self._stop_requested.is_set():
+                    break
+                if self._reset_requested.is_set():
+                    self._reset_requested.clear()
+                    self._reset_robot()
+                    continue
+                if self._start_requested.is_set():
+                    self._start_requested.clear()
+                    self._run_segment()
+                    continue
+                self._wake.wait(timeout=self._POLL_INTERVAL_S)
+                self._wake.clear()
+        finally:
+            self._listener.stop()
+            log_say("Interactive session ended", play_sounds)
+
+    def _run_segment(self) -> None:
+        """Execute one ``strategy.run`` segment until interrupted or finished."""
+        engine = self._ctx.policy.inference
+        # Clear the local flag *before* checking the request flags: command
+        # handlers set their flag first and the segment-stop event second, so
+        # a /reset or /stop racing with this /start is either seen here or
+        # ends the freshly started loop on its first tick.
+        self._segment_stop.clear()
+        if self._stop_requested.is_set() or self._reset_requested.is_set() or self._global_shutdown.is_set():
+            return
+        self._strategy.reset_control_state()
+        log_say("Starting rollout", self._ctx.runtime.cfg.play_sounds)
+        self._print("Rollout running — /reset to pause and return to initial position, /stop to shut down.")
+        self._running.set()
+        try:
+            self._strategy.run(self._ctx)
+        finally:
+            self._running.clear()
+            engine.pause()
+        if engine.failed:
+            return  # the session loop reports the failure and shuts down
+        if not (
+            self._stop_requested.is_set() or self._reset_requested.is_set() or self._global_shutdown.is_set()
+        ):
+            self._print(
+                "Rollout run ended on its own (duration reached). Robot is holding position — "
+                "/start to run again, /reset to return to initial position, /stop to shut down."
+            )
+
+    def _reset_robot(self) -> None:
+        """Pause inference and return the robot to its initial position."""
+        self._ctx.policy.inference.pause()
+        log_say("Resetting robot to initial position", self._ctx.runtime.cfg.play_sounds)
+        if self._ctx.hardware.initial_position:
+            self._strategy.return_to_initial_position(self._ctx.hardware)
+            self._print("Robot reset — holding at initial position. /start to run.")
+        else:
+            logger.warning("No initial position captured — skipping the return move")
+            self._print("Robot paused — no initial position captured, holding current pose. /start to run.")
+
+    # ------------------------------------------------------------------
+    # Command handlers (called from the listener thread; only set flags)
+    # ------------------------------------------------------------------
+
+    def _handle_line(self, line: str) -> None:
+        cmd = parse_command(line)
+        if cmd is None:
+            self._print("Input not recognized — commands start with '/'. Type /help for the list.")
+            return
+        entry = self._commands.get(cmd.name)
+        if entry is None:
+            self._print(f"Unknown command '/{cmd.name}'. Type /help for the list.")
+            return
+        handler, _ = entry
+        handler(cmd)
+
+    def _handle_eof(self) -> None:
+        self._print("Input stream closed — stopping the session.")
+        self._request_stop()
+
+    def _cmd_start(self, cmd: InteractiveCommand) -> None:
+        if self._running.is_set():
+            self._print("Already running — /reset to pause first, or /stop to shut down.")
+            return
+        self._start_requested.set()
+        self._wake.set()
+
+    def _cmd_reset(self, cmd: InteractiveCommand) -> None:
+        # Last command wins: a /start still waiting to be serviced is cancelled
+        # so the robot never starts moving after the operator asked it not to.
+        # Flag first, segment-stop second (see the ordering note in _run_segment).
+        self._start_requested.clear()
+        self._reset_requested.set()
+        self._segment_stop.set()
+        self._wake.set()
+
+    def _cmd_stop(self, cmd: InteractiveCommand) -> None:
+        self._request_stop()
+
+    def _request_stop(self) -> None:
+        self._start_requested.clear()  # last command wins, see _cmd_reset
+        self._stop_requested.set()
+        self._segment_stop.set()
+        self._wake.set()
+
+    def _cmd_help(self, cmd: InteractiveCommand) -> None:
+        self._print(self._render_help())
+
+    # ------------------------------------------------------------------
+    # Rendering
+    # ------------------------------------------------------------------
+
+    def _render_help(self) -> str:
+        width = max(len(name) for name in self._commands)
+        lines = [f"  /{name:<{width}}   {help_line}" for name, (_, help_line) in self._commands.items()]
+        return "Available commands:\n" + "\n".join(lines)
+
+    def _render_banner(self) -> str:
+        return (
+            f"{_BANNER_RULE}\n"
+            "Interactive rollout session — the robot will NOT move until you type /start.\n"
+            f"{self._render_help()}\n"
+            f"{_BANNER_RULE}"
+        )
+
+    @staticmethod
+    def _print(message: str) -> None:
+        """User-facing chat output; logging stays on stderr, replies on stdout."""
+        print(message, flush=True)

src/lerobot/rollout/strategies/core.py

@@ -63,12 +63,25 @@ class RolloutStrategy(abc.ABC):
         self._interpolator = ActionInterpolator(multiplier=ctx.runtime.cfg.interpolation_multiplier)
         self._engine = ctx.policy.inference
         logger.info("Starting inference engine...")
-        self._engine.reset()
+        self.reset_control_state()
         self._engine.start()
         self._warmup_flushed = False
-        self._cached_obs_processed = None
         logger.info("Inference engine started")
 
+    def reset_control_state(self) -> None:
+        """Clear episode-scoped control state so a paused session can restart cleanly.
+
+        Resets the inference engine (policy hidden state, action queues), the
+        action interpolator, and the cached processed observation.  Used by the
+        interactive session between run segments; only call while the control
+        loop is not running.
+        """
+        if self._engine is not None:
+            self._engine.reset()
+        if self._interpolator is not None:
+            self._interpolator.reset()
+        self._cached_obs_processed = None
+
     def _process_observation_and_notify(self, processors: ProcessorContext, obs_raw: dict) -> dict:
         """Run the observation processor and notify the engine — throttled to policy ticks.
 
@@ -125,7 +138,7 @@ class RolloutStrategy(abc.ABC):
         if robot.is_connected:
             if return_to_initial_position and hw.initial_position:
                 logger.info("Returning robot to initial position before shutdown...")
-                self._return_to_initial_position(hw)
+                self.return_to_initial_position(hw)
             elif not return_to_initial_position:
                 logger.info(
                     "Skipping return-to-initial-position (disabled by config); leaving robot in final pose."
@@ -138,7 +151,7 @@ class RolloutStrategy(abc.ABC):
             teleop.disconnect()
 
     @staticmethod
-    def _return_to_initial_position(hw: HardwareContext, duration_s: float = 3.0, fps: int = 50) -> None:
+    def return_to_initial_position(hw: HardwareContext, duration_s: float = 3.0, fps: int = 50) -> None:
         """Smoothly interpolate the robot back to its initial position."""
         robot = hw.robot_wrapper
         target = hw.initial_position

src/lerobot/rollout/strategies/episodic.py

@@ -165,7 +165,7 @@ class EpisodicStrategy(RolloutStrategy):
 
                         elif self.config.reset_to_initial_position:
                             # No teleop: return the robot to its startup position.
-                            self._return_to_initial_position(hw=ctx.hardware, duration_s=1)
+                            self.return_to_initial_position(hw=ctx.hardware, duration_s=1)
 
                         self._reset_loop(
                             ctx=ctx,
@@ -187,7 +187,7 @@ class EpisodicStrategy(RolloutStrategy):
 
                         # returns to its initial joint positions captured at startup
                         if not teleop and self.config.reset_to_initial_position:
-                            self._return_to_initial_position(hw=ctx.hardware, duration_s=1)
+                            self.return_to_initial_position(hw=ctx.hardware, duration_s=1)
 
                         continue
 

src/lerobot/scripts/lerobot_rollout.py

@@ -44,6 +44,17 @@ Usage examples
         --robot.port=/dev/ttyACM0 \\
         --task="pick up cube" --duration=30
 
+    # Base mode — interactive session: the robot stays idle until /start is
+    # typed; /reset returns it to the initial position (hardware and policy
+    # stay warm); /stop shuts down gracefully
+    lerobot-rollout \\
+        --strategy.type=base \\
+        --policy.path=lerobot/act_koch_real \\
+        --robot.type=koch_follower \\
+        --robot.port=/dev/ttyACM0 \\
+        --task="pick up cube" \\
+        --interactive=true
+
     # Base mode — RTC inference for slow VLAs (Pi0, Pi0.5, SmolVLA)
     lerobot-rollout \\
         --strategy.type=base \\
@@ -173,7 +184,13 @@ from lerobot.robots import (  # noqa: F401
     so_follower,
     unitree_g1 as unitree_g1_robot,
 )
-from lerobot.rollout import RolloutConfig, build_rollout_context, create_strategy
+from lerobot.rollout import (
+    InteractiveSession,
+    LinkedEvent,
+    RolloutConfig,
+    build_rollout_context,
+    create_strategy,
+)
 from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
@@ -215,6 +232,10 @@ def rollout(cfg: RolloutConfig):
 
     signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
     shutdown_event = signal_handler.shutdown_event
+    if cfg.interactive:
+        # Session commands (/reset, /stop) end the running control loop by setting
+        # the local flag; process signals still propagate through the parent event.
+        shutdown_event = LinkedEvent(shutdown_event)
 
     logger.info("Building rollout context...")
     ctx = build_rollout_context(cfg, shutdown_event)
@@ -230,8 +251,12 @@ def rollout(cfg: RolloutConfig):
 
     try:
         strategy.setup(ctx)
-        logger.info("Rollout setup complete, starting rollout...")
+        if cfg.interactive:
-        strategy.run(ctx)
+            logger.info("Rollout setup complete — starting interactive session (robot idle until /start)")
+            InteractiveSession(strategy, ctx).run()
+        else:
+            logger.info("Rollout setup complete, starting rollout...")
+            strategy.run(ctx)
     except KeyboardInterrupt:
         logger.info("Interrupted by user")
     finally:

tests/test_interactive_rollout.py

@@ -0,0 +1,529 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for the interactive rollout session (--interactive=true)."""
+
+from __future__ import annotations
+
+import contextlib
+import os
+import time
+from threading import Event, Thread
+from types import SimpleNamespace
+from unittest.mock import MagicMock
+
+import pytest
+
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+
+from lerobot.rollout import (  # noqa: E402
+    InteractiveCommand,
+    InteractiveSession,
+    LinkedEvent,
+    StdinCommandListener,
+    parse_command,
+)
+
+
+def _wait_for(predicate, timeout: float = 2.0) -> bool:
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        if predicate():
+            return True
+        time.sleep(0.005)
+    return predicate()
+
+
+@contextlib.contextmanager
+def _pipe_stream():
+    """A held-open pipe so the session's stdin listener never sees EOF."""
+    read_fd, write_fd = os.pipe()
+    reader = os.fdopen(read_fd, "r")
+    writer = os.fdopen(write_fd, "w")
+    try:
+        yield reader, writer
+    finally:
+        with contextlib.suppress(OSError, ValueError):
+            writer.close()
+        with contextlib.suppress(OSError, ValueError):
+            reader.close()
+
+
+# ---------------------------------------------------------------------------
+# Command parser
+# ---------------------------------------------------------------------------
+
+
+def test_parse_command_basic():
+    cmd = parse_command("/start")
+    assert cmd is not None
+    assert cmd.name == "start"
+    assert cmd.args == ""
+
+
+def test_parse_command_case_whitespace_and_args():
+    cmd = parse_command("  /SubTask Grab the red cube  ")
+    assert cmd is not None
+    assert cmd.name == "subtask"
+    assert cmd.args == "Grab the red cube"
+
+
+def test_parse_command_tab_separated():
+    assert parse_command("/subtask\tgrab the cube") == InteractiveCommand(
+        name="subtask", args="grab the cube"
+    )
+
+
+def test_parse_command_non_commands():
+    assert parse_command("hello robot") is None
+    assert parse_command("") is None
+    assert parse_command("   ") is None
+    assert parse_command("/") is None
+    assert parse_command("/ start") is None
+
+
+# ---------------------------------------------------------------------------
+# LinkedEvent
+# ---------------------------------------------------------------------------
+
+
+def test_linked_event_local_flag():
+    parent = Event()
+    event = LinkedEvent(parent)
+    assert not event.is_set()
+
+    event.set()
+    assert event.is_set()
+    assert not parent.is_set()
+
+    event.clear()
+    assert not event.is_set()
+
+
+def test_linked_event_reflects_parent():
+    parent = Event()
+    event = LinkedEvent(parent)
+    parent.set()
+    assert event.is_set()
+    # Clearing the local flag never masks the parent.
+    event.clear()
+    assert event.is_set()
+
+
+def test_linked_event_wait():
+    parent = Event()
+    event = LinkedEvent(parent)
+    assert event.wait(timeout=0.05) is False
+
+    parent.set()
+    assert event.wait(timeout=0.05) is True
+
+    parent.clear()
+    event.set()
+    assert event.wait(timeout=0.05) is True
+
+
+def test_linked_event_wait_wakes_on_parent_set():
+    parent = Event()
+    event = LinkedEvent(parent)
+    Thread(target=lambda: (time.sleep(0.05), parent.set()), daemon=True).start()
+    assert event.wait(timeout=2.0) is True
+
+
+# ---------------------------------------------------------------------------
+# StdinCommandListener
+# ---------------------------------------------------------------------------
+
+
+def test_stdin_listener_reads_lines_and_eof():
+    lines: list[str] = []
+    eof = Event()
+
+    with _pipe_stream() as (reader, writer):
+        listener = StdinCommandListener(lines.append, on_eof=eof.set, stream=reader)
+        listener.start()
+        writer.write("/start\n")
+        writer.write("   \n")  # blank lines are skipped
+        writer.write("/help\n")
+        writer.flush()
+        assert _wait_for(lambda: len(lines) == 2)
+        assert lines == ["/start", "/help"]
+
+        writer.close()
+        assert _wait_for(eof.is_set)
+        listener.stop()
+
+
+def test_stdin_listener_handler_errors_do_not_kill_reader():
+    lines: list[str] = []
+
+    def flaky(line: str) -> None:
+        if line == "/boom":
+            raise RuntimeError("boom")
+        lines.append(line)
+
+    with _pipe_stream() as (reader, writer):
+        listener = StdinCommandListener(flaky, stream=reader)
+        listener.start()
+        writer.write("/boom\n/start\n")
+        writer.flush()
+        assert _wait_for(lambda: lines == ["/start"])
+        listener.stop()
+
+
+def test_stdin_listener_blocking_fallback():
+    """Streams without a file descriptor (e.g. StringIO) use the blocking readline path."""
+    import io
+
+    lines: list[str] = []
+    eof = Event()
+    listener = StdinCommandListener(lines.append, on_eof=eof.set, stream=io.StringIO("/start\n\n/help\n"))
+    assert not listener._use_select
+    listener.start()
+    assert _wait_for(eof.is_set)
+    assert lines == ["/start", "/help"]
+    listener.stop()
+
+
+# ---------------------------------------------------------------------------
+# InteractiveSession
+# ---------------------------------------------------------------------------
+
+
+def _make_session(input_stream, run_behavior=None):
+    """Build a session around a mock strategy and a minimal fake context."""
+    parent = Event()
+    stop_event = LinkedEvent(parent)
+    engine = MagicMock()
+    engine.failed = False
+    ctx = SimpleNamespace(
+        runtime=SimpleNamespace(
+            cfg=SimpleNamespace(play_sounds=False),
+            shutdown_event=stop_event,
+        ),
+        policy=SimpleNamespace(inference=engine),
+        hardware=SimpleNamespace(initial_position={"joint.pos": 0.0}),
+    )
+
+    strategy = MagicMock()
+    run_started = Event()
+
+    def default_run(c):
+        run_started.set()
+        while not c.runtime.shutdown_event.is_set():
+            time.sleep(0.005)
+
+    strategy.run.side_effect = run_behavior or default_run
+
+    session = InteractiveSession(strategy, ctx, input_stream=input_stream)
+    return session, strategy, engine, parent, run_started
+
+
+def _start_session_thread(session) -> Thread:
+    thread = Thread(target=session.run, daemon=True)
+    thread.start()
+    return thread
+
+
+def test_session_requires_linked_event():
+    ctx = SimpleNamespace(runtime=SimpleNamespace(shutdown_event=Event()))
+    with pytest.raises(TypeError, match="LinkedEvent"):
+        InteractiveSession(MagicMock(), ctx)
+
+
+def test_session_start_reset_restart_stop_flow():
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, engine, _parent, run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+
+        # Idle until /start: the strategy loop must not run on its own.
+        time.sleep(0.05)
+        strategy.run.assert_not_called()
+
+        session._handle_line("/start")
+        assert _wait_for(run_started.is_set)
+        assert strategy.reset_control_state.call_count == 1
+
+        # /reset ends the segment, pauses the engine, and returns to the initial position.
+        session._handle_line("/reset")
+        assert _wait_for(lambda: strategy.return_to_initial_position.call_count == 1)
+        assert engine.pause.call_count >= 1
+        assert thread.is_alive()
+
+        # /start again runs a fresh segment with freshly reset control state.
+        run_started.clear()
+        session._handle_line("/start")
+        assert _wait_for(run_started.is_set)
+        assert strategy.run.call_count == 2
+        assert strategy.reset_control_state.call_count == 2
+
+        # /stop ends the session; teardown stays with the caller (the CLI script).
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+        strategy.teardown.assert_not_called()
+
+
+def test_session_stop_while_idle():
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+        strategy.run.assert_not_called()
+
+
+def test_session_reset_while_idle_returns_to_initial_position():
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+        session._handle_line("/reset")
+        assert _wait_for(lambda: strategy.return_to_initial_position.call_count == 1)
+        assert thread.is_alive()
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+
+
+def test_session_start_while_running_is_rejected():
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+
+        session._handle_line("/start")
+        assert _wait_for(run_started.is_set)
+        session._handle_line("/start")
+        time.sleep(0.05)
+        assert strategy.run.call_count == 1
+
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+
+
+def test_session_reset_cancels_pending_start():
+    """Last command wins: a queued /start must not fire after a later /reset."""
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader)
+        # Queue both commands before the session loop starts servicing them.
+        session._handle_line("/start")
+        session._handle_line("/reset")
+        thread = _start_session_thread(session)
+
+        assert _wait_for(lambda: strategy.return_to_initial_position.call_count == 1)
+        time.sleep(0.05)
+        strategy.run.assert_not_called()
+
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+
+
+def test_session_stop_cancels_pending_start():
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader)
+        session._handle_line("/start")
+        session._handle_line("/stop")
+        thread = _start_session_thread(session)
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+        strategy.run.assert_not_called()
+
+
+def test_session_exits_on_parent_shutdown():
+    with _pipe_stream() as (reader, _writer):
+        session, _strategy, _engine, parent, run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+        session._handle_line("/start")
+        assert _wait_for(run_started.is_set)
+
+        parent.set()  # SIGINT/SIGTERM path
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+
+
+def test_session_stops_on_engine_failure():
+    def failing_run(c):
+        # Mimic the RTC thread's fatal-error path: flag the failure and set
+        # the shutdown event the engine was built with (the LinkedEvent).
+        c.policy.inference.failed = True
+        c.runtime.shutdown_event.set()
+
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader, run_behavior=failing_run)
+        thread = _start_session_thread(session)
+        session._handle_line("/start")
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+        # A failed engine ends the session instead of returning to idle.
+        strategy.return_to_initial_position.assert_not_called()
+
+
+def test_session_stops_on_engine_failure_while_idle():
+    """A fatal engine error while idle ends the session instead of being masked by /start."""
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, engine, _parent, _run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+        time.sleep(0.05)
+        engine.failed = True
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+        strategy.run.assert_not_called()
+
+
+def test_session_returns_to_idle_when_run_ends_naturally():
+    def finite_run(c):
+        return None  # e.g. --duration elapsed
+
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader, run_behavior=finite_run)
+        thread = _start_session_thread(session)
+
+        session._handle_line("/start")
+        assert _wait_for(lambda: strategy.run.call_count == 1)
+        time.sleep(0.05)
+        assert thread.is_alive()  # back to idle, not shut down
+
+        # The session accepts another /start after a natural end.
+        session._handle_line("/start")
+        assert _wait_for(lambda: strategy.run.call_count == 2)
+
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+
+
+def test_session_unknown_input_does_not_start(capsys):
+    with _pipe_stream() as (reader, _writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+
+        session._handle_line("/frobnicate")
+        session._handle_line("hello robot")
+        session._handle_line("/help")
+        time.sleep(0.05)
+        strategy.run.assert_not_called()
+
+        out = capsys.readouterr().out
+        assert "/frobnicate" in out
+        assert "commands start with '/'" in out
+
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+
+
+def test_session_eof_stops_session():
+    with _pipe_stream() as (reader, writer):
+        session, strategy, _engine, _parent, _run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+        writer.close()  # EOF on the command stream
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+        strategy.run.assert_not_called()
+
+
+def test_session_commands_via_stream():
+    """End-to-end: commands flow through the pipe and the listener thread."""
+    with _pipe_stream() as (reader, writer):
+        session, strategy, _engine, _parent, run_started = _make_session(reader)
+        thread = _start_session_thread(session)
+
+        writer.write("/start\n")
+        writer.flush()
+        assert _wait_for(run_started.is_set)
+
+        writer.write("/stop\n")
+        writer.flush()
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+        assert strategy.run.call_count == 1
+
+
+def test_session_drives_real_base_strategy():
+    """End-to-end with a real BaseStrategy control loop (only hardware/engine mocked)."""
+    from lerobot.rollout import BaseStrategy, BaseStrategyConfig
+
+    parent = Event()
+    stop_event = LinkedEvent(parent)
+    engine = MagicMock()
+    engine.failed = False
+    engine.get_action.return_value = None  # no action ready; the loop still ticks
+
+    robot = MagicMock()
+    robot.get_observation.return_value = {"joint.pos": 0.0}
+
+    def identity(x):
+        return x
+
+    ctx = SimpleNamespace(
+        runtime=SimpleNamespace(
+            cfg=SimpleNamespace(
+                play_sounds=False,
+                fps=100.0,
+                duration=0.0,
+                use_torch_compile=False,
+                interpolation_multiplier=1,
+                display_data=False,
+            ),
+            shutdown_event=stop_event,
+        ),
+        policy=SimpleNamespace(inference=engine),
+        hardware=SimpleNamespace(robot_wrapper=robot, teleop=None, initial_position={"joint.pos": 0.0}),
+        processors=SimpleNamespace(
+            teleop_action_processor=identity,
+            robot_action_processor=identity,
+            robot_observation_processor=identity,
+        ),
+        data=SimpleNamespace(dataset=None, dataset_features={}, hw_features={}, ordered_action_keys=[]),
+    )
+
+    strategy = BaseStrategy(BaseStrategyConfig())
+    strategy.setup(ctx)
+    strategy.return_to_initial_position = MagicMock()  # skip the 3s hardware sweep
+
+    with _pipe_stream() as (reader, _writer):
+        session = InteractiveSession(strategy, ctx, input_stream=reader)
+        thread = _start_session_thread(session)
+
+        session._handle_line("/start")
+        assert _wait_for(lambda: engine.resume.called)
+        assert _wait_for(lambda: robot.get_observation.call_count >= 3)
+
+        session._handle_line("/reset")
+        assert _wait_for(lambda: strategy.return_to_initial_position.called)
+        assert engine.pause.called
+        assert thread.is_alive()
+
+        session._handle_line("/start")
+        assert _wait_for(lambda: engine.resume.call_count >= 2)
+
+        session._handle_line("/stop")
+        thread.join(timeout=2.0)
+        assert not thread.is_alive()
+
+
+# ---------------------------------------------------------------------------
+# Config validation
+# ---------------------------------------------------------------------------
+
+
+def test_interactive_requires_base_strategy():
+    from lerobot.configs.dataset import DatasetRecordConfig
+    from lerobot.rollout import RolloutConfig, SentryStrategyConfig
+    from tests.mocks.mock_robot import MockRobotConfig
+
+    with pytest.raises(ValueError, match="--interactive=true currently supports only"):
+        RolloutConfig(
+            robot=MockRobotConfig(),
+            strategy=SentryStrategyConfig(),
+            dataset=DatasetRecordConfig(repo_id="user/rollout_test", single_task="test"),
+            interactive=True,
+        )