src/lerobot/rl/actor.py

#!/usr/bin/env python

# Copyright 2025 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.
"""

Actor server runner for distributed HILSerl robot policy training.



This script implements the actor component of the distributed HILSerl architecture.

It executes the policy in the robot environment, collects experience,

and sends transitions to the learner server for policy updates.



Examples of usage:



- Start an actor server for real robot training with human-in-the-loop intervention:

```bash

python -m lerobot.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json

```



**NOTE**: The actor server requires a running learner server to connect to. Ensure the learner

server is started before launching the actor.



**NOTE**: Human intervention is key to HILSerl training. Press the upper right trigger button on the

gamepad to take control of the robot during training. Initially intervene frequently, then gradually

reduce interventions as the policy improves.



**WORKFLOW**:

1. Determine robot workspace bounds using `lerobot-find-joint-limits`

2. Record demonstrations with `gym_manipulator.py` in record mode

3. Process the dataset and determine camera crops with `crop_dataset_roi.py`

4. Start the learner server with the training configuration

5. Start this actor server with the same configuration

6. Use human interventions to guide policy learning



For more details on the complete HILSerl training workflow, see:

https://github.com/michel-aractingi/lerobot-hilserl-guide

"""

import logging
import os
import time
from functools import lru_cache
from queue import Empty

import grpc
import torch
from torch import nn
from torch.multiprocessing import Event, Queue

from lerobot.cameras import opencv  # noqa: F401
from lerobot.configs import parser
from lerobot.configs.train import TrainRLServerPipelineConfig
from lerobot.policies.factory import make_policy
from lerobot.policies.sac.modeling_sac import SACPolicy
from lerobot.processor import TransitionKey
from lerobot.rl.process import ProcessSignalHandler
from lerobot.rl.queue import get_last_item_from_queue
from lerobot.robots import so100_follower  # noqa: F401
from lerobot.teleoperators import gamepad, so101_leader  # noqa: F401
from lerobot.teleoperators.utils import TeleopEvents
from lerobot.transport import services_pb2, services_pb2_grpc
from lerobot.transport.utils import (
    bytes_to_state_dict,
    grpc_channel_options,
    python_object_to_bytes,
    receive_bytes_in_chunks,
    send_bytes_in_chunks,
    transitions_to_bytes,
)
from lerobot.utils.random_utils import set_seed
from lerobot.utils.robot_utils import busy_wait
from lerobot.utils.transition import (
    Transition,
    move_state_dict_to_device,
    move_transition_to_device,
)
from lerobot.utils.utils import (
    TimerManager,
    get_safe_torch_device,
    init_logging,
)

from .gym_manipulator import (
    create_transition,
    make_processors,
    make_robot_env,
    step_env_and_process_transition,
)

# Main entry point


@parser.wrap()
def actor_cli(cfg: TrainRLServerPipelineConfig):
    cfg.validate()
    display_pid = False
    if not use_threads(cfg):
        import torch.multiprocessing as mp

        mp.set_start_method("spawn")
        display_pid = True

    # Create logs directory to ensure it exists
    log_dir = os.path.join(cfg.output_dir, "logs")
    os.makedirs(log_dir, exist_ok=True)
    log_file = os.path.join(log_dir, f"actor_{cfg.job_name}.log")

    # Initialize logging with explicit log file
    init_logging(log_file=log_file, display_pid=display_pid)
    logging.info(f"Actor logging initialized, writing to {log_file}")

    is_threaded = use_threads(cfg)
    shutdown_event = ProcessSignalHandler(is_threaded, display_pid=display_pid).shutdown_event

    learner_client, grpc_channel = learner_service_client(
        host=cfg.policy.actor_learner_config.learner_host,
        port=cfg.policy.actor_learner_config.learner_port,
    )

    logging.info("[ACTOR] Establishing connection with Learner")
    if not establish_learner_connection(learner_client, shutdown_event):
        logging.error("[ACTOR] Failed to establish connection with Learner")
        return

    if not use_threads(cfg):
        # If we use multithreading, we can reuse the channel
        grpc_channel.close()
        grpc_channel = None

    logging.info("[ACTOR] Connection with Learner established")

    parameters_queue = Queue()
    transitions_queue = Queue()
    interactions_queue = Queue()

    concurrency_entity = None
    if use_threads(cfg):
        from threading import Thread

        concurrency_entity = Thread
    else:
        from multiprocessing import Process

        concurrency_entity = Process

    receive_policy_process = concurrency_entity(
        target=receive_policy,
        args=(cfg, parameters_queue, shutdown_event, grpc_channel),
        daemon=True,
    )

    transitions_process = concurrency_entity(
        target=send_transitions,
        args=(cfg, transitions_queue, shutdown_event, grpc_channel),
        daemon=True,
    )

    interactions_process = concurrency_entity(
        target=send_interactions,
        args=(cfg, interactions_queue, shutdown_event, grpc_channel),
        daemon=True,
    )

    transitions_process.start()
    interactions_process.start()
    receive_policy_process.start()

    act_with_policy(
        cfg=cfg,
        shutdown_event=shutdown_event,
        parameters_queue=parameters_queue,
        transitions_queue=transitions_queue,
        interactions_queue=interactions_queue,
    )
    logging.info("[ACTOR] Policy process joined")

    logging.info("[ACTOR] Closing queues")
    transitions_queue.close()
    interactions_queue.close()
    parameters_queue.close()

    transitions_process.join()
    logging.info("[ACTOR] Transitions process joined")
    interactions_process.join()
    logging.info("[ACTOR] Interactions process joined")
    receive_policy_process.join()
    logging.info("[ACTOR] Receive policy process joined")

    logging.info("[ACTOR] join queues")
    transitions_queue.cancel_join_thread()
    interactions_queue.cancel_join_thread()
    parameters_queue.cancel_join_thread()

    logging.info("[ACTOR] queues closed")


# Core algorithm functions


def act_with_policy(

    cfg: TrainRLServerPipelineConfig,

    shutdown_event: any,  # Event,

    parameters_queue: Queue,

    transitions_queue: Queue,

    interactions_queue: Queue,

):
    """

    Executes policy interaction within the environment.



    This function rolls out the policy in the environment, collecting interaction data and pushing it to a queue for streaming to the learner.

    Once an episode is completed, updated network parameters received from the learner are retrieved from a queue and loaded into the network.



    Args:

        cfg: Configuration settings for the interaction process.

        shutdown_event: Event to check if the process should shutdown.

        parameters_queue: Queue to receive updated network parameters from the learner.

        transitions_queue: Queue to send transitions to the learner.

        interactions_queue: Queue to send interactions to the learner.

    """
    # Initialize logging for multiprocessing
    if not use_threads(cfg):
        log_dir = os.path.join(cfg.output_dir, "logs")
        os.makedirs(log_dir, exist_ok=True)
        log_file = os.path.join(log_dir, f"actor_policy_{os.getpid()}.log")
        init_logging(log_file=log_file, display_pid=True)
        logging.info("Actor policy process logging initialized")

    logging.info("make_env online")

    online_env, teleop_device = make_robot_env(cfg=cfg.env)
    env_processor, action_processor = make_processors(online_env, teleop_device, cfg.env, cfg.policy.device)

    set_seed(cfg.seed)
    device = get_safe_torch_device(cfg.policy.device, log=True)

    torch.backends.cudnn.benchmark = True
    torch.backends.cuda.matmul.allow_tf32 = True

    logging.info("make_policy")

    ### Instantiate the policy in both the actor and learner processes
    ### To avoid sending a SACPolicy object through the port, we create a policy instance
    ### on both sides, the learner sends the updated parameters every n steps to update the actor's parameters
    policy: SACPolicy = make_policy(
        cfg=cfg.policy,
        env_cfg=cfg.env,
    )
    policy = policy.eval()
    assert isinstance(policy, nn.Module)

    obs, info = online_env.reset()
    env_processor.reset()
    action_processor.reset()

    # Process initial observation
    transition = create_transition(observation=obs, info=info)
    transition = env_processor(transition)

    # NOTE: For the moment we will solely handle the case of a single environment
    sum_reward_episode = 0
    list_transition_to_send_to_learner = []
    episode_intervention = False
    # Add counters for intervention rate calculation
    episode_intervention_steps = 0
    episode_total_steps = 0

    policy_timer = TimerManager("Policy inference", log=False)

    for interaction_step in range(cfg.policy.online_steps):
        start_time = time.perf_counter()
        if shutdown_event.is_set():
            logging.info("[ACTOR] Shutting down act_with_policy")
            return

        observation = {
            k: v for k, v in transition[TransitionKey.OBSERVATION].items() if k in cfg.policy.input_features
        }

        # Time policy inference and check if it meets FPS requirement
        with policy_timer:
            # Extract observation from transition for policy
            action = policy.select_action(batch=observation)
        policy_fps = policy_timer.fps_last

        log_policy_frequency_issue(policy_fps=policy_fps, cfg=cfg, interaction_step=interaction_step)

        # Use the new step function
        new_transition = step_env_and_process_transition(
            env=online_env,
            transition=transition,
            action=action,
            env_processor=env_processor,
            action_processor=action_processor,
        )

        # Extract values from processed transition
        next_observation = {
            k: v
            for k, v in new_transition[TransitionKey.OBSERVATION].items()
            if k in cfg.policy.input_features
        }

        # Teleop action is the action that was executed in the environment
        # It is either the action from the teleop device or the action from the policy
        executed_action = new_transition[TransitionKey.COMPLEMENTARY_DATA]["teleop_action"]

        reward = new_transition[TransitionKey.REWARD]
        done = new_transition.get(TransitionKey.DONE, False)
        truncated = new_transition.get(TransitionKey.TRUNCATED, False)

        sum_reward_episode += float(reward)
        episode_total_steps += 1

        # Check for intervention from transition info
        intervention_info = new_transition[TransitionKey.INFO]
        if intervention_info.get(TeleopEvents.IS_INTERVENTION, False):
            episode_intervention = True
            episode_intervention_steps += 1

        complementary_info = {
            "discrete_penalty": torch.tensor(
                [new_transition[TransitionKey.COMPLEMENTARY_DATA].get("discrete_penalty", 0.0)]
            ),
        }
        # Create transition for learner (convert to old format)
        list_transition_to_send_to_learner.append(
            Transition(
                state=observation,
                action=executed_action,
                reward=reward,
                next_state=next_observation,
                done=done,
                truncated=truncated,
                complementary_info=complementary_info,
            )
        )

        # Update transition for next iteration
        transition = new_transition

        if done or truncated:
            logging.info(f"[ACTOR] Global step {interaction_step}: Episode reward: {sum_reward_episode}")

            update_policy_parameters(policy=policy, parameters_queue=parameters_queue, device=device)

            if len(list_transition_to_send_to_learner) > 0:
                push_transitions_to_transport_queue(
                    transitions=list_transition_to_send_to_learner,
                    transitions_queue=transitions_queue,
                )
                list_transition_to_send_to_learner = []

            stats = get_frequency_stats(policy_timer)
            policy_timer.reset()

            # Calculate intervention rate
            intervention_rate = 0.0
            if episode_total_steps > 0:
                intervention_rate = episode_intervention_steps / episode_total_steps

            # Send episodic reward to the learner
            interactions_queue.put(
                python_object_to_bytes(
                    {
                        "Episodic reward": sum_reward_episode,
                        "Interaction step": interaction_step,
                        "Episode intervention": int(episode_intervention),
                        "Intervention rate": intervention_rate,
                        **stats,
                    }
                )
            )

            # Reset intervention counters and environment
            sum_reward_episode = 0.0
            episode_intervention = False
            episode_intervention_steps = 0
            episode_total_steps = 0

            # Reset environment and processors
            obs, info = online_env.reset()
            env_processor.reset()
            action_processor.reset()

            # Process initial observation
            transition = create_transition(observation=obs, info=info)
            transition = env_processor(transition)

        if cfg.env.fps is not None:
            dt_time = time.perf_counter() - start_time
            busy_wait(1 / cfg.env.fps - dt_time)


#  Communication Functions - Group all gRPC/messaging functions


def establish_learner_connection(

    stub: services_pb2_grpc.LearnerServiceStub,

    shutdown_event: Event,  # type: ignore

    attempts: int = 30,

):
    """Establish a connection with the learner.



    Args:

        stub (services_pb2_grpc.LearnerServiceStub): The stub to use for the connection.

        shutdown_event (Event): The event to check if the connection should be established.

        attempts (int): The number of attempts to establish the connection.

    Returns:

        bool: True if the connection is established, False otherwise.

    """
    for _ in range(attempts):
        if shutdown_event.is_set():
            logging.info("[ACTOR] Shutting down establish_learner_connection")
            return False

        # Force a connection attempt and check state
        try:
            logging.info("[ACTOR] Send ready message to Learner")
            if stub.Ready(services_pb2.Empty()) == services_pb2.Empty():
                return True
        except grpc.RpcError as e:
            logging.error(f"[ACTOR] Waiting for Learner to be ready... {e}")
            time.sleep(2)
    return False


@lru_cache(maxsize=1)
def learner_service_client(

    host: str = "127.0.0.1",

    port: int = 50051,

) -> tuple[services_pb2_grpc.LearnerServiceStub, grpc.Channel]:
    """

    Returns a client for the learner service.



    GRPC uses HTTP/2, which is a binary protocol and multiplexes requests over a single connection.

    So we need to create only one client and reuse it.

    """

    channel = grpc.insecure_channel(
        f"{host}:{port}",
        grpc_channel_options(),
    )
    stub = services_pb2_grpc.LearnerServiceStub(channel)
    logging.info("[ACTOR] Learner service client created")
    return stub, channel


def receive_policy(

    cfg: TrainRLServerPipelineConfig,

    parameters_queue: Queue,

    shutdown_event: Event,  # type: ignore

    learner_client: services_pb2_grpc.LearnerServiceStub | None = None,

    grpc_channel: grpc.Channel | None = None,

):
    """Receive parameters from the learner.



    Args:

        cfg (TrainRLServerPipelineConfig): The configuration for the actor.

        parameters_queue (Queue): The queue to receive the parameters.

        shutdown_event (Event): The event to check if the process should shutdown.

    """
    logging.info("[ACTOR] Start receiving parameters from the Learner")
    if not use_threads(cfg):
        # Create a process-specific log file
        log_dir = os.path.join(cfg.output_dir, "logs")
        os.makedirs(log_dir, exist_ok=True)
        log_file = os.path.join(log_dir, f"actor_receive_policy_{os.getpid()}.log")

        # Initialize logging with explicit log file
        init_logging(log_file=log_file, display_pid=True)
        logging.info("Actor receive policy process logging initialized")

        # Setup process handlers to handle shutdown signal
        # But use shutdown event from the main process
        _ = ProcessSignalHandler(use_threads=False, display_pid=True)

    if grpc_channel is None or learner_client is None:
        learner_client, grpc_channel = learner_service_client(
            host=cfg.policy.actor_learner_config.learner_host,
            port=cfg.policy.actor_learner_config.learner_port,
        )

    try:
        iterator = learner_client.StreamParameters(services_pb2.Empty())
        receive_bytes_in_chunks(
            iterator,
            parameters_queue,
            shutdown_event,
            log_prefix="[ACTOR] parameters",
        )

    except grpc.RpcError as e:
        logging.error(f"[ACTOR] gRPC error: {e}")

    if not use_threads(cfg):
        grpc_channel.close()
    logging.info("[ACTOR] Received policy loop stopped")


def send_transitions(

    cfg: TrainRLServerPipelineConfig,

    transitions_queue: Queue,

    shutdown_event: any,  # Event,

    learner_client: services_pb2_grpc.LearnerServiceStub | None = None,

    grpc_channel: grpc.Channel | None = None,

) -> services_pb2.Empty:
    """

    Sends transitions to the learner.



    This function continuously retrieves messages from the queue and processes:



    - Transition Data:

        - A batch of transitions (observation, action, reward, next observation) is collected.

        - Transitions are moved to the CPU and serialized using PyTorch.

        - The serialized data is wrapped in a `services_pb2.Transition` message and sent to the learner.

    """

    if not use_threads(cfg):
        # Create a process-specific log file
        log_dir = os.path.join(cfg.output_dir, "logs")
        os.makedirs(log_dir, exist_ok=True)
        log_file = os.path.join(log_dir, f"actor_transitions_{os.getpid()}.log")

        # Initialize logging with explicit log file
        init_logging(log_file=log_file, display_pid=True)
        logging.info("Actor transitions process logging initialized")

    if grpc_channel is None or learner_client is None:
        learner_client, grpc_channel = learner_service_client(
            host=cfg.policy.actor_learner_config.learner_host,
            port=cfg.policy.actor_learner_config.learner_port,
        )

    try:
        learner_client.SendTransitions(
            transitions_stream(
                shutdown_event, transitions_queue, cfg.policy.actor_learner_config.queue_get_timeout
            )
        )
    except grpc.RpcError as e:
        logging.error(f"[ACTOR] gRPC error: {e}")

    logging.info("[ACTOR] Finished streaming transitions")

    if not use_threads(cfg):
        grpc_channel.close()
    logging.info("[ACTOR] Transitions process stopped")


def send_interactions(

    cfg: TrainRLServerPipelineConfig,

    interactions_queue: Queue,

    shutdown_event: Event,  # type: ignore

    learner_client: services_pb2_grpc.LearnerServiceStub | None = None,

    grpc_channel: grpc.Channel | None = None,

) -> services_pb2.Empty:
    """

    Sends interactions to the learner.



    This function continuously retrieves messages from the queue and processes:



    - Interaction Messages:

        - Contains useful statistics about episodic rewards and policy timings.

        - The message is serialized using `pickle` and sent to the learner.

    """

    if not use_threads(cfg):
        # Create a process-specific log file
        log_dir = os.path.join(cfg.output_dir, "logs")
        os.makedirs(log_dir, exist_ok=True)
        log_file = os.path.join(log_dir, f"actor_interactions_{os.getpid()}.log")

        # Initialize logging with explicit log file
        init_logging(log_file=log_file, display_pid=True)
        logging.info("Actor interactions process logging initialized")

        # Setup process handlers to handle shutdown signal
        # But use shutdown event from the main process
        _ = ProcessSignalHandler(use_threads=False, display_pid=True)

    if grpc_channel is None or learner_client is None:
        learner_client, grpc_channel = learner_service_client(
            host=cfg.policy.actor_learner_config.learner_host,
            port=cfg.policy.actor_learner_config.learner_port,
        )

    try:
        learner_client.SendInteractions(
            interactions_stream(
                shutdown_event, interactions_queue, cfg.policy.actor_learner_config.queue_get_timeout
            )
        )
    except grpc.RpcError as e:
        logging.error(f"[ACTOR] gRPC error: {e}")

    logging.info("[ACTOR] Finished streaming interactions")

    if not use_threads(cfg):
        grpc_channel.close()
    logging.info("[ACTOR] Interactions process stopped")


def transitions_stream(shutdown_event: Event, transitions_queue: Queue, timeout: float) -> services_pb2.Empty:  # type: ignore
    while not shutdown_event.is_set():
        try:
            message = transitions_queue.get(block=True, timeout=timeout)
        except Empty:
            logging.debug("[ACTOR] Transition queue is empty")
            continue

        yield from send_bytes_in_chunks(
            message, services_pb2.Transition, log_prefix="[ACTOR] Send transitions"
        )

    return services_pb2.Empty()


def interactions_stream(

    shutdown_event: Event,

    interactions_queue: Queue,

    timeout: float,  # type: ignore

) -> services_pb2.Empty:
    while not shutdown_event.is_set():
        try:
            message = interactions_queue.get(block=True, timeout=timeout)
        except Empty:
            logging.debug("[ACTOR] Interaction queue is empty")
            continue

        yield from send_bytes_in_chunks(
            message,
            services_pb2.InteractionMessage,
            log_prefix="[ACTOR] Send interactions",
        )

    return services_pb2.Empty()


#  Policy functions


def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device):
    bytes_state_dict = get_last_item_from_queue(parameters_queue, block=False)
    if bytes_state_dict is not None:
        logging.info("[ACTOR] Load new parameters from Learner.")
        state_dicts = bytes_to_state_dict(bytes_state_dict)

        # TODO: check encoder parameter synchronization possible issues:
        # 1. When shared_encoder=True, we're loading stale encoder params from actor's state_dict
        #    instead of the updated encoder params from critic (which is optimized separately)
        # 2. When freeze_vision_encoder=True, we waste bandwidth sending/loading frozen params
        # 3. Need to handle encoder params correctly for both actor and discrete_critic
        # Potential fixes:
        # - Send critic's encoder state when shared_encoder=True
        # - Skip encoder params entirely when freeze_vision_encoder=True
        # - Ensure discrete_critic gets correct encoder state (currently uses encoder_critic)

        # Load actor state dict
        actor_state_dict = move_state_dict_to_device(state_dicts["policy"], device=device)
        policy.actor.load_state_dict(actor_state_dict)

        # Load discrete critic if present
        if hasattr(policy, "discrete_critic") and "discrete_critic" in state_dicts:
            discrete_critic_state_dict = move_state_dict_to_device(
                state_dicts["discrete_critic"], device=device
            )
            policy.discrete_critic.load_state_dict(discrete_critic_state_dict)
            logging.info("[ACTOR] Loaded discrete critic parameters from Learner.")


#  Utilities functions


def push_transitions_to_transport_queue(transitions: list, transitions_queue):
    """Send transitions to learner in smaller chunks to avoid network issues.



    Args:

        transitions: List of transitions to send

        message_queue: Queue to send messages to learner

        chunk_size: Size of each chunk to send

    """
    transition_to_send_to_learner = []
    for transition in transitions:
        tr = move_transition_to_device(transition=transition, device="cpu")
        for key, value in tr["state"].items():
            if torch.isnan(value).any():
                logging.warning(f"Found NaN values in transition {key}")

        transition_to_send_to_learner.append(tr)

    transitions_queue.put(transitions_to_bytes(transition_to_send_to_learner))


def get_frequency_stats(timer: TimerManager) -> dict[str, float]:
    """Get the frequency statistics of the policy.



    Args:

        timer (TimerManager): The timer with collected metrics.



    Returns:

        dict[str, float]: The frequency statistics of the policy.

    """
    stats = {}
    if timer.count > 1:
        avg_fps = timer.fps_avg
        p90_fps = timer.fps_percentile(90)
        logging.debug(f"[ACTOR] Average policy frame rate: {avg_fps}")
        logging.debug(f"[ACTOR] Policy frame rate 90th percentile: {p90_fps}")
        stats = {
            "Policy frequency [Hz]": avg_fps,
            "Policy frequency 90th-p [Hz]": p90_fps,
        }
    return stats


def log_policy_frequency_issue(policy_fps: float, cfg: TrainRLServerPipelineConfig, interaction_step: int):
    if policy_fps < cfg.env.fps:
        logging.warning(
            f"[ACTOR] Policy FPS {policy_fps:.1f} below required {cfg.env.fps} at step {interaction_step}"
        )


def use_threads(cfg: TrainRLServerPipelineConfig) -> bool:
    return cfg.policy.concurrency.actor == "threads"


if __name__ == "__main__":
    actor_cli()