blanchon's picture
Update
3380376
import asyncio
import contextlib
import logging
import time
from collections import deque
import numpy as np
from transport_server_client import RoboticsConsumer, RoboticsProducer
from transport_server_client.video import VideoConsumer, VideoProducer
from inference_server.models import get_inference_engine
from inference_server.models.joint_config import JointConfig
logger = logging.getLogger(__name__)
def busy_wait(seconds):
"""
Precise timing function for consistent control loops.
On some systems, asyncio.sleep is not accurate enough for
control loops, so we use busy waiting for short delays.
"""
if seconds > 0:
end_time = asyncio.get_event_loop().time() + seconds
while asyncio.get_event_loop().time() < end_time:
pass
class InferenceSession:
"""
A single inference session managing one model and its Transport Server connections.
Handles joint values in NORMALIZED VALUES throughout the pipeline.
Supports multiple camera streams with different camera names.
Supports multiple policy types: ACT, Pi0, SmolVLA, Diffusion Policy.
"""
def __init__(
self,
session_id: str,
policy_path: str,
camera_names: list[str],
transport_server_url: str,
workspace_id: str,
camera_room_ids: dict[str, str],
joint_input_room_id: str,
joint_output_room_id: str,
policy_type: str = "act",
language_instruction: str | None = None,
):
self.session_id = session_id
self.policy_path = policy_path
self.policy_type = policy_type.lower()
self.camera_names = camera_names
self.transport_server_url = transport_server_url
self.language_instruction = language_instruction
# Workspace and Room IDs
self.workspace_id = workspace_id
self.camera_room_ids = camera_room_ids # {camera_name: room_id}
self.joint_input_room_id = joint_input_room_id
self.joint_output_room_id = joint_output_room_id
# Transport clients - multiple camera consumers
self.camera_consumers: dict[str, VideoConsumer] = {} # {camera_name: consumer}
self.joint_input_consumer: RoboticsConsumer | None = None
self.joint_output_producer: RoboticsProducer | None = None
# Generic inference engine (supports all policy types)
self.inference_engine = None
# Session state
self.status = "initializing"
self.error_message: str | None = None
self.inference_task: asyncio.Task | None = None
# Data buffers - all in normalized values
self.latest_images: dict[str, np.ndarray] = {} # {camera_name: image}
self.latest_joint_positions: np.ndarray | None = None
# Complete joint state (always 6 joints) - initialized with zeros
self.complete_joint_state: np.ndarray = np.zeros(6, dtype=np.float32)
self.images_updated: dict[str, bool] = dict.fromkeys(camera_names, False)
self.joints_updated = False
# Action queue for proper chunking (important for ACT, optional for others)
self.action_queue: deque = deque(maxlen=100) # Adjust maxlen as needed
self.n_action_steps = 10 # How many actions to use from each chunk
# Memory optimization: Clear old actions periodically
self.last_queue_cleanup = time.time()
self.queue_cleanup_interval = 10.0 # seconds
# Control frequency configuration
self.control_frequency_hz = 20 # Hz - reduced from 30 to improve performance
self.inference_frequency_hz = 2 # Hz - reduced from 3 to improve performance
# Statistics
self.stats = {
"inference_count": 0,
"images_received": dict.fromkeys(camera_names, 0),
"joints_received": 0,
"commands_sent": 0,
"errors": 0,
"actions_in_queue": 0,
"policy_type": self.policy_type,
}
# Robot responsiveness tracking
self.last_command_values: np.ndarray | None = None
self.last_joint_check_time = time.time()
# Session timeout management
self.last_activity_time = time.time()
self.timeout_seconds = 600 # 10 minutes
self.timeout_check_task: asyncio.Task | None = None
async def initialize(self):
"""Initialize the session by loading the model and setting up Transport Server connections."""
logger.info(
f"Initializing session {self.session_id} with policy type: {self.policy_type}, "
f"cameras: {self.camera_names}"
)
# Initialize inference engine based on policy type
engine_kwargs = {
"policy_path": self.policy_path,
"camera_names": self.camera_names,
}
# Add language instruction for policies that support it
if (
self.policy_type in {"pi0", "pi0fast", "smolvla"}
and self.language_instruction
):
engine_kwargs["language_instruction"] = self.language_instruction
self.inference_engine = get_inference_engine(self.policy_type, **engine_kwargs)
# Load the policy
await self.inference_engine.load_policy()
# Create Transport clients for each camera
for camera_name in self.camera_names:
self.camera_consumers[camera_name] = VideoConsumer(
self.transport_server_url
)
self.joint_input_consumer = RoboticsConsumer(self.transport_server_url)
self.joint_output_producer = RoboticsProducer(self.transport_server_url)
# Set up callbacks
self._setup_callbacks()
# Connect to rooms
await self._connect_to_rooms()
# Start receiving video frames from all cameras
for camera_name, consumer in self.camera_consumers.items():
await consumer.start_receiving()
logger.info(f"Started receiving frames for camera: {camera_name}")
# Start timeout monitoring
self.timeout_check_task = asyncio.create_task(self._timeout_monitor())
self.status = "ready"
logger.info(
f"✅ Session {self.session_id} initialized successfully with {self.policy_type} policy "
f"and {len(self.camera_names)} cameras"
)
def _setup_callbacks(self):
"""Set up callbacks for Transport clients."""
def create_frame_callback(camera_name: str):
"""Create a frame callback for a specific camera."""
def on_frame_received(frame_data):
"""Handle incoming camera frame from VideoConsumer."""
metadata = frame_data.metadata
width = metadata.get("width", 0)
height = metadata.get("height", 0)
format_type = metadata.get("format", "rgb24")
if format_type == "rgb24" and width > 0 and height > 0:
# Convert bytes to numpy array (server sends RGB format)
frame_bytes = frame_data.data
# Validate frame data size
expected_size = height * width * 3
if len(frame_bytes) != expected_size:
logger.warning(
f"Frame size mismatch for camera {camera_name}: "
f"expected {expected_size}, got {len(frame_bytes)}. Skipping frame."
)
self.stats["errors"] += 1
return
img_rgb = np.frombuffer(frame_bytes, dtype=np.uint8).reshape((
height,
width,
3,
))
# Store as latest image for inference
self.latest_images[camera_name] = img_rgb
self.images_updated[camera_name] = True
self.stats["images_received"][camera_name] += 1
# Update activity time
self.last_activity_time = time.time()
return on_frame_received
# Set up frame callbacks for each camera
for camera_name in self.camera_names:
callback = create_frame_callback(camera_name)
self.camera_consumers[camera_name].on_frame_update(callback)
def on_joints_received(joints_data):
"""Handle incoming joint data from RoboticsConsumer."""
joint_values = self._parse_joint_data(joints_data)
if joint_values:
# Update complete joint state with received values
for i, value in enumerate(joint_values[:6]): # Ensure max 6 joints
self.complete_joint_state[i] = value
self.latest_joint_positions = self.complete_joint_state.copy()
self.joints_updated = True
self.stats["joints_received"] += 1
# Update activity time
self.last_activity_time = time.time()
def on_error(error_msg):
"""Handle Transport client errors."""
logger.error(
f"Transport client error in session {self.session_id}: {error_msg}"
)
self.error_message = str(error_msg)
self.stats["errors"] += 1
# Set callbacks for RoboticsConsumer
self.joint_input_consumer.on_joint_update(on_joints_received)
self.joint_input_consumer.on_state_sync(on_joints_received)
self.joint_input_consumer.on_error(on_error)
# Set error callbacks for VideoConsumers
for consumer in self.camera_consumers.values():
consumer.on_error(on_error)
def _parse_joint_data(self, joints_data) -> list[float]:
"""
Parse joint data from Transport Server message.
Args:
joints_data: Joint data from Transport Server message
Returns:
List of 6 normalized joint values in standard order
"""
return JointConfig.parse_joint_data(joints_data, self.policy_type)
async def _connect_to_rooms(self):
"""Connect to all Transport Server rooms."""
# Connect to camera rooms as consumer
for camera_name, consumer in self.camera_consumers.items():
room_id = self.camera_room_ids[camera_name]
success = await consumer.connect(
self.workspace_id, room_id, f"{self.session_id}-{camera_name}-consumer"
)
if not success:
msg = f"Failed to connect to camera room for {camera_name}"
logger.error(msg)
logger.info(
f"Connected to camera room for {camera_name}: {room_id} in workspace {self.workspace_id}"
)
# Connect to joint input room as consumer
success = await self.joint_input_consumer.connect(
self.workspace_id,
self.joint_input_room_id,
f"{self.session_id}-joint-input-consumer",
)
if not success:
msg = "Failed to connect to joint input room"
logger.error(msg)
# Connect to joint output room as producer
success = await self.joint_output_producer.connect(
self.workspace_id,
self.joint_output_room_id,
f"{self.session_id}-joint-output-producer",
)
if not success:
msg = "Failed to connect to joint output room"
logger.error(msg)
logger.info(
f"Connected to all rooms for session {self.session_id} in workspace {self.workspace_id}"
)
async def start_inference(self):
"""Start the inference loop."""
if self.status != "ready":
msg = f"Session not ready. Current status: {self.status}"
logger.error(msg)
self.status = "running"
self.inference_task = asyncio.create_task(self._inference_loop())
logger.info(f"Started inference for session {self.session_id}")
async def stop_inference(self):
"""Stop the inference loop."""
if self.inference_task:
self.inference_task.cancel()
with contextlib.suppress(asyncio.CancelledError):
await self.inference_task
self.inference_task = None
self.status = "stopped"
logger.info(f"Stopped inference for session {self.session_id}")
async def restart_inference(self):
"""Restart the inference loop (stop if running, then start)."""
logger.info(f"Restarting inference for session {self.session_id}")
# Stop current inference if running
await self.stop_inference()
# Reset internal state for fresh start
self._reset_session_state()
# Start inference again
await self.start_inference()
logger.info(f"Successfully restarted inference for session {self.session_id}")
def _reset_session_state(self):
"""Reset session state for restart."""
# Clear action queue
self.action_queue.clear()
# Reset complete joint state to zeros
self.complete_joint_state.fill(0.0)
# Reset image update flags
for camera_name in self.camera_names:
self.images_updated[camera_name] = False
self.joints_updated = False
# Reset timing
self.last_queue_cleanup = time.time()
# Reset inference engine state if available
if self.inference_engine:
self.inference_engine.reset()
# Reset some statistics (but keep cumulative counts)
self.stats["actions_in_queue"] = 0
logger.info(f"Reset session state for {self.session_id}")
async def _timeout_monitor(self):
"""Monitor session for inactivity timeout."""
while True:
try:
await asyncio.sleep(60) # Check every minute
current_time = time.time()
inactive_time = current_time - self.last_activity_time
if inactive_time > self.timeout_seconds:
logger.warning(
f"Session {self.session_id} has been inactive for "
f"{inactive_time:.1f} seconds (timeout: {self.timeout_seconds}s). "
f"Marking for cleanup."
)
self.status = "timeout"
# Don't cleanup here - let the session manager handle it
break
if inactive_time > self.timeout_seconds * 0.8: # Warn at 80% of timeout
logger.info(
f"Session {self.session_id} inactive for {inactive_time:.1f}s, "
f"will timeout in {self.timeout_seconds - inactive_time:.1f}s"
)
except asyncio.CancelledError:
logger.info(f"Timeout monitor cancelled for session {self.session_id}")
break
except Exception:
logger.exception(
f"Error in timeout monitor for session {self.session_id}"
)
break
def _all_cameras_have_data(self) -> bool:
"""Check if we have received data from all cameras."""
return all(
camera_name in self.latest_images for camera_name in self.camera_names
)
async def _inference_loop(self):
"""Main inference loop that processes incoming data and sends commands."""
logger.info(f"Starting inference loop for session {self.session_id}")
logger.info(
f"Control frequency: {self.control_frequency_hz} Hz, Inference frequency: {self.inference_frequency_hz} Hz"
)
logger.info(
f"Waiting for data from {len(self.camera_names)} cameras: {self.camera_names}"
)
inference_counter = 0
target_dt = 1.0 / self.control_frequency_hz # Control loop period
inference_interval = (
self.control_frequency_hz // self.inference_frequency_hz
) # How many control steps per inference
while True:
loop_start_time = asyncio.get_event_loop().time()
# Check if we have images from all cameras and joint data
if (
self._all_cameras_have_data()
and self.latest_joint_positions is not None
):
# Removed verbose data logging
# Only run inference at the specified frequency and when queue is empty
should_run_inference = len(self.action_queue) == 0 or (
inference_counter % inference_interval == 0
and len(self.action_queue) < 3
)
if should_run_inference:
# Only log inference runs occasionally to reduce overhead
if self.stats["inference_count"] % 10 == 0:
logger.info(
f"Running inference #{self.stats['inference_count']} for session {self.session_id} "
f"(queue length: {len(self.action_queue)})"
)
# Verify joint positions have correct shape before inference
if self.latest_joint_positions.shape != (6,):
logger.error(
f"Invalid joint positions shape: {self.latest_joint_positions.shape}, "
f"expected (6,). Values: {self.latest_joint_positions}"
)
# Fix the shape by resetting to complete joint state
self.latest_joint_positions = self.complete_joint_state.copy()
# Prepare inference arguments
inference_kwargs = {
"images": self.latest_images,
"joint_positions": self.latest_joint_positions,
}
# Add language instruction for vision-language policies
if (
self.policy_type in {"pi0", "pi0fast", "smolvla"}
and self.language_instruction
):
inference_kwargs["task"] = self.language_instruction
# Run inference to get action chunk
predicted_actions = await self.inference_engine.predict(
**inference_kwargs
)
# ACT returns a chunk of actions, we need to queue them
if len(predicted_actions.shape) == 1:
# Single action returned, use it directly
actions_to_queue = [predicted_actions]
else:
# Multiple actions in chunk, take first n_action_steps
actions_to_queue = predicted_actions[: self.n_action_steps]
# Add actions to queue
for action in actions_to_queue:
joint_commands = (
self.inference_engine.get_joint_commands_with_names(action)
)
self.action_queue.append(joint_commands)
self.stats["inference_count"] += 1
# Reset image update flags
for camera_name in self.camera_names:
self.images_updated[camera_name] = False
self.joints_updated = False
# Send action from queue if available
if len(self.action_queue) > 0:
joint_commands = self.action_queue.popleft()
# Only log commands occasionally
if self.stats["commands_sent"] % 100 == 0:
logger.info(
f"🤖 Sent {self.stats['commands_sent']} commands. Latest: {joint_commands[0]['name']}={joint_commands[0]['value']:.1f}"
)
await self.joint_output_producer.send_joint_update(joint_commands)
self.stats["commands_sent"] += 1
self.stats["actions_in_queue"] = len(self.action_queue)
# Store command values for responsiveness check
command_values = np.array(
[cmd["value"] for cmd in joint_commands], dtype=np.float32
)
self.last_command_values = command_values
# Periodic memory cleanup
current_time = asyncio.get_event_loop().time()
if current_time - self.last_queue_cleanup > self.queue_cleanup_interval:
# Clear stale actions if queue is getting full
if len(self.action_queue) > 80: # 80% of maxlen
self.action_queue.clear()
self.last_queue_cleanup = current_time
inference_counter += 1
# Precise timing control for consistent control frequency
elapsed_time = asyncio.get_event_loop().time() - loop_start_time
sleep_time = target_dt - elapsed_time
if sleep_time > 0.001: # Use asyncio.sleep for longer waits
await asyncio.sleep(sleep_time)
elif sleep_time > 0: # Use busy_wait for precise short delays
busy_wait(sleep_time)
elif sleep_time < -0.01: # Log if we're running significantly slow
logger.warning(
f"Control loop running slow for session {self.session_id}: {elapsed_time * 1000:.1f}ms (target: {target_dt * 1000:.1f}ms)"
)
async def cleanup(self):
"""Clean up session resources."""
logger.info(f"Cleaning up session {self.session_id}")
# Stop timeout monitoring
if self.timeout_check_task:
self.timeout_check_task.cancel()
with contextlib.suppress(asyncio.CancelledError):
await self.timeout_check_task
self.timeout_check_task = None
# Stop inference
await self.stop_inference()
# Disconnect Transport clients
for camera_name, consumer in self.camera_consumers.items():
await consumer.stop_receiving()
await consumer.disconnect()
logger.info(f"Disconnected camera consumer for {camera_name}")
if self.joint_input_consumer:
await self.joint_input_consumer.disconnect()
if self.joint_output_producer:
await self.joint_output_producer.disconnect()
# Clean up inference engine
if self.inference_engine:
del self.inference_engine
self.inference_engine = None
logger.info(f"Session {self.session_id} cleanup completed")
def get_status(self) -> dict:
"""Get current session status."""
status_dict = {
"session_id": self.session_id,
"status": self.status,
"policy_path": self.policy_path,
"policy_type": self.policy_type,
"camera_names": self.camera_names,
"workspace_id": self.workspace_id,
"rooms": {
"workspace_id": self.workspace_id,
"camera_room_ids": self.camera_room_ids,
"joint_input_room_id": self.joint_input_room_id,
"joint_output_room_id": self.joint_output_room_id,
},
"stats": self.stats.copy(),
"error_message": self.error_message,
"joint_state": {
"complete_joint_state": self.complete_joint_state.tolist(),
"latest_joint_positions": (
self.latest_joint_positions.tolist()
if self.latest_joint_positions is not None
else None
),
"joint_state_shape": (
self.latest_joint_positions.shape
if self.latest_joint_positions is not None
else None
),
},
}
return status_dict
class SessionManager:
"""Manages multiple inference sessions and their lifecycle."""
def __init__(self):
self.sessions: dict[str, InferenceSession] = {}
self.cleanup_task: asyncio.Task | None = None
self._start_cleanup_task()
def _start_cleanup_task(self):
"""Start the automatic cleanup task for timed-out sessions."""
try:
# Only start if we're in an async context
loop = asyncio.get_running_loop()
self.cleanup_task = loop.create_task(self._periodic_cleanup())
except RuntimeError:
# No event loop running yet, will start later
pass
async def _periodic_cleanup(self):
"""Periodically check for and clean up timed-out sessions."""
while True:
try:
await asyncio.sleep(300) # Check every 5 minutes
# Find sessions that have timed out
timed_out_sessions = []
for session_id, session in self.sessions.items():
if session.status == "timeout":
timed_out_sessions.append(session_id)
# Clean up timed-out sessions
for session_id in timed_out_sessions:
logger.info(f"Auto-cleaning up timed-out session: {session_id}")
await self.delete_session(session_id)
except asyncio.CancelledError:
logger.info("Periodic cleanup task cancelled")
break
async def create_session(
self,
session_id: str,
policy_path: str,
transport_server_url: str,
camera_names: list[str] | None = None,
workspace_id: str | None = None,
policy_type: str = "act",
language_instruction: str | None = None,
) -> dict[str, str]:
"""Create a new inference session."""
if camera_names is None:
camera_names = ["front"]
if session_id in self.sessions:
msg = f"Session {session_id} already exists"
raise ValueError(msg)
# Create camera rooms using VideoProducer
video_temp_client = VideoProducer(transport_server_url)
camera_room_ids = {}
# Use provided workspace_id or create new one
if workspace_id:
target_workspace_id = workspace_id
logger.info(
f"Using provided workspace ID {target_workspace_id} for session {session_id}"
)
# Create all camera rooms in the specified workspace
for camera_name in camera_names:
_, room_id = await video_temp_client.create_room(
workspace_id=target_workspace_id,
room_id=f"{session_id}-{camera_name}",
)
camera_room_ids[camera_name] = room_id
else:
# Create first camera room to get new workspace_id
first_camera = camera_names[0]
target_workspace_id, first_room_id = await video_temp_client.create_room(
room_id=f"{session_id}-{first_camera}"
)
logger.info(
f"Generated new workspace ID {target_workspace_id} for session {session_id}"
)
# Store the first room
camera_room_ids[first_camera] = first_room_id
# Create remaining camera rooms in the same workspace
for camera_name in camera_names[1:]:
_, room_id = await video_temp_client.create_room(
workspace_id=target_workspace_id,
room_id=f"{session_id}-{camera_name}",
)
camera_room_ids[camera_name] = room_id
# Create joint rooms using RoboticsProducer in the same workspace
robotics_temp_client = RoboticsProducer(transport_server_url)
_, joint_input_room_id = await robotics_temp_client.create_room(
workspace_id=target_workspace_id, room_id=f"{session_id}-joint-input"
)
_, joint_output_room_id = await robotics_temp_client.create_room(
workspace_id=target_workspace_id, room_id=f"{session_id}-joint-output"
)
logger.info(
f"Created rooms for session {session_id} in workspace {target_workspace_id}:"
)
for camera_name, room_id in camera_room_ids.items():
logger.info(f" Camera room ({camera_name}): {room_id}")
logger.info(f" Joint input room: {joint_input_room_id}")
logger.info(f" Joint output room: {joint_output_room_id}")
# Create session
session = InferenceSession(
session_id=session_id,
policy_path=policy_path,
camera_names=camera_names,
transport_server_url=transport_server_url,
workspace_id=target_workspace_id,
camera_room_ids=camera_room_ids,
joint_input_room_id=joint_input_room_id,
joint_output_room_id=joint_output_room_id,
policy_type=policy_type,
language_instruction=language_instruction,
)
# Initialize session
await session.initialize()
# Store session
self.sessions[session_id] = session
# Start cleanup task if not already running
if not self.cleanup_task or self.cleanup_task.done():
self._start_cleanup_task()
return {
"workspace_id": target_workspace_id,
"camera_room_ids": camera_room_ids,
"joint_input_room_id": joint_input_room_id,
"joint_output_room_id": joint_output_room_id,
}
async def start_inference(self, session_id: str):
"""Start inference for a specific session."""
if session_id not in self.sessions:
msg = f"Session {session_id} not found"
raise KeyError(msg)
await self.sessions[session_id].start_inference()
async def stop_inference(self, session_id: str):
"""Stop inference for a specific session."""
if session_id not in self.sessions:
msg = f"Session {session_id} not found"
raise KeyError(msg)
await self.sessions[session_id].stop_inference()
async def restart_inference(self, session_id: str):
"""Restart inference for a specific session."""
if session_id not in self.sessions:
msg = f"Session {session_id} not found"
raise KeyError(msg)
await self.sessions[session_id].restart_inference()
async def delete_session(self, session_id: str):
"""Delete a session and clean up all resources."""
if session_id not in self.sessions:
msg = f"Session {session_id} not found"
raise KeyError(msg)
session = self.sessions[session_id]
await session.cleanup()
del self.sessions[session_id]
logger.info(f"Deleted session {session_id}")
async def list_sessions(self) -> list[dict]:
"""List all sessions with their status."""
return [session.get_status() for session in self.sessions.values()]
async def cleanup_all_sessions(self):
"""Clean up all sessions."""
logger.info("Cleaning up all sessions...")
# Stop the cleanup task
if self.cleanup_task:
self.cleanup_task.cancel()
with contextlib.suppress(asyncio.CancelledError):
await self.cleanup_task
self.cleanup_task = None
# Clean up all sessions
for session_id in list(self.sessions.keys()):
await self.delete_session(session_id)
logger.info("All sessions cleaned up")