finish load inference server model

This commit is contained in:
hufei.hofee
2026-03-18 18:12:34 +08:00
commit cc9815f3b8
4 changed files with 476 additions and 0 deletions

1
.gitignore vendored Normal file
View File

@@ -0,0 +1 @@
.vscode

179
benchmark.yaml Normal file
View File

@@ -0,0 +1,179 @@
general:
scan_project: true
root_paths:
asset: /home/ubuntu/projects/gen_data/data
output: /home/ubuntu/output
checkpoints: /home/ubuntu/data/models
simulation:
launch_config:
device: cuda
enable_cameras: true
headless: false
livestream: 0
scene:
name: default_scene_name
position: [0, 0, 0]
rotation: [1, 0, 0, 0]
base_config:
name: default_base
source: primitive
stereotype: ground_plane
ground_size: [100,100]
object_cfg_dict:
table:
name: simple_table
position: [0.5, 0, 0.25]
source: primitive
stereotype: rigid
primitive_type: cuboid
primitive_size: [0.5, 1, 0.5]
mass: 1e4
target:
name: target
position: [0.4, 0.0, 0.5]
scale: [0.001, 0.001, 0.001]
axis_y_up: true
asset_path: asset://objects/omni6DPose/ball/omni6DPose_ball_020/Aligned.usd
stereotype: rigid
source: local
robot_cfg_dict:
robot:
name: my_robot
asset_path: asset://Franka/franka_robotiq_2f85_zedmini.usd
position: [0, 0, 0]
stereotype: single_gripper_arm_robot
source: local
init_joint_position:
panda_joint2: -0.1633
panda_joint4: -1.070
panda_joint6: 0.8933
panda_joint7: 0.785
arm_actuator_name: franka_arm
gripper_actuator_name: robotiq_2f_85
use_planner: true
planner_cfg:
stereotype: curobo
lazy_init: true
robot_config_file: asset://curobo/franka_robotiq_2f85/franka_robotiq_2f85.yml
world_config_source: stage
world_stage_ignore_substrings: [my_robot]
world_stage_only_paths: [/World]
world_stage_reference_prim_path: /World/Robot/SingleGripperArmRobot/my_robot
sensor_cfg_dict:
front_camera:
name: front_camera
stereotype: camera
position: [0.8, 0.0, 0.8]
data_types: [rgb, depth, normals]
width: 1280
height: 720
camera_model: pinhole
fix_camera: true
left_camera:
name: left_camera
stereotype: camera
position: [0.6, 0.7, 0.8]
data_types: [rgb, depth, normals]
width: 1280
height: 720
camera_model: pinhole
fix_camera: true
right_camera:
name: right_camera
stereotype: camera
position: [0.6, -0.7, 0.8]
data_types: [rgb, depth, normals]
width: 1280
height: 720
camera_model: pinhole
fix_camera: true
extension:
extension_cfg_dict:
my_data_collect:
enable: true
stereotype: data_collect
observer_cfgs:
- stereotype: robot_observer
name: my_robot
observe_joint_positions: true
observe_joint_velocities: true
observe_joint_accelerations: true
observe_joint_position_targets: true
observe_joint_velocity_targets: true
observe_position: true
observe_rotation: true
observe_ee_pose: true
observe_gripper_state: true
observe_gripper_drive_state: true
- stereotype: sensor_observer
name: front_camera
observe_intrinsic_matrix: true
observe_extrinsic_matrix: true
observe_rgb: true
observe_depth: true
observe_normals: true
- stereotype: sensor_observer
name: left_camera
observe_intrinsic_matrix: true
observe_extrinsic_matrix: true
observe_rgb: true
observe_depth: true
observe_normals: true
- stereotype: sensor_observer
name: right_camera
observe_intrinsic_matrix: true
observe_extrinsic_matrix: true
observe_rgb: true
observe_depth: true
observe_normals: true
- stereotype: task_observer
name: task
- stereotype: object_observer
name: target
observe_position: true
observe_rotation: true
observe_scale: true
my_benchmark:
enable: true
stereotype: benchmark
data_collector_name: my_data_collect
goals:
- name: reach_target
description: Reach the target
stereotype: pose
pose_A_source: ee
pose_A_params:
robot_name: my_robot
pose_B_source: spawnable
pose_B_params:
spawnable_name: target
position_tolerance: 0.005
policy:
stereotype: starvla
robot_name: my_robot
object_name: target
prompt: pick the cola bottle and place it on the book
policy_server:
ckpt_path: checkpoints://0309_qwenpi_droid_cola_post/final_model/pytorch_model.pt
ckpt_source: local
host: 0.0.0.0
port: 5000
use_bf16: true
unnorm_key: oxe_bridge
state_mode: ee_pose7

173
starvla_inference_server.py Normal file
View File

@@ -0,0 +1,173 @@
import yaml
import pickle
import os
from urllib.parse import urlparse
import numpy as np
import torch
import cv2
from flask import Flask, request, Response
from PIL import Image
class StarvlaInferenceServer:
def __init__(self, config_path: str):
with open(config_path, "r") as f:
cfg = yaml.safe_load(f)
policy_server_cfg = cfg["policy_server"]
root_paths = cfg["general"]["root_paths"]
self.ckpt_source = policy_server_cfg["ckpt_source"]
self.ckpt_path = self._resolve_ckpt_path(
ckpt_url=policy_server_cfg["ckpt_path"],
root_paths=root_paths,
)
self.host = policy_server_cfg.get("host", "0.0.0.0")
self.port = policy_server_cfg.get("port", 5000)
self.use_bf16 = policy_server_cfg.get("use_bf16", True)
self.unnorm_key = policy_server_cfg.get("unnorm_key", "oxe_bridge")
self.state_mode = policy_server_cfg.get("state_mode", "ee_pose7")
print("Loading StarVLA model...")
self.model = self.load_model()
print("Model loaded.")
self.app = Flask(__name__)
self.register_routes()
@staticmethod
def _resolve_ckpt_path(ckpt_url: str, root_paths: dict) -> str:
parsed = urlparse(ckpt_url)
if not parsed.scheme:
return ckpt_url
root = root_paths.get(parsed.scheme)
if not root:
raise KeyError(
f"cannot find the checkpoint root path in root_paths: {root_paths}"
)
rel = (parsed.netloc + parsed.path).lstrip("/")
return os.path.join(root, rel)
def load_model(self):
from starVLA.model.framework.share_tools import read_mode_config, dict_to_namespace
from starVLA.model.framework.__init__ import build_framework
model_config, norm_stats = read_mode_config(self.ckpt_path)
cfg = dict_to_namespace(model_config)
cfg.trainer.pretrained_checkpoint = None
model = build_framework(cfg=cfg)
model.norm_stats = norm_stats
state_dict = torch.load(self.ckpt_path, map_location="cpu")
model.load_state_dict(state_dict, strict=True)
if self.use_bf16:
model = model.to(torch.bfloat16)
model = model.to("cuda").eval()
self.norm_stats = norm_stats
self.action_norm_stats = norm_stats.get(self.unnorm_key, {}).get("action", None)
return model
def parse_observation(self, obs):
rgb = obs["rgb"][-1]
state = obs["state"][-1]
joint = obs.get("joint", None)
prompt = obs["prompt"]
left = rgb[:, :, :3]
right = rgb[:, :, 3:6]
wrist = rgb[:, :, 6:9]
target_size = (320, 180)
left = cv2.resize(left, target_size)
right = cv2.resize(right, target_size)
wrist = cv2.resize(wrist, target_size)
img_left = Image.fromarray(left)
img_right = Image.fromarray(right)
img_wrist = Image.fromarray(wrist)
if self.state_mode == "joint8":
joint_last = joint[-1]
gripper = state[9]
state_vec = np.concatenate(
[joint_last, np.array([gripper])],
axis=0
)
else:
xyz = state[0:3]
rot6d = state[3:9]
gripper = state[9]
state_vec = np.concatenate(
[xyz, rot6d[:3], np.array([gripper])],
axis=0
)
return img_left, img_right, img_wrist, state_vec, prompt
def inference(self, observation: dict) -> dict:
img_left, img_right, img_wrist, state_vec, prompt = \
self.parse_observation(observation)
vla_input = {
"batch_images": [[img_left, img_right, img_wrist]],
"instructions": [prompt],
"state": [state_vec]
}
with torch.no_grad():
output = self.model.predict_action(**vla_input)
actions = output.get("normalized_actions")
if isinstance(actions, torch.Tensor):
actions = actions.cpu().numpy()
if actions.ndim == 3:
actions = actions[0]
return {"action": actions.astype(np.float32)}
def register_routes(self):
@self.app.route("/policy", methods=["POST"])
def policy():
data = pickle.loads(request.data)
result = self.inference(data)
body = pickle.dumps(result, protocol=4)
return Response(body, mimetype="application/octet-stream")
def run(self):
print("StarVLA policy server running")
print(f"Host: {self.host}")
print(f"Port: {self.port}")
self.app.run(
host=self.host,
port=self.port,
threaded=True
)
if __name__ == "__main__":
config_path = "./benchmark.yaml"
server = StarvlaInferenceServer(config_path)
server.run()

123
starvla_policy.py Normal file
View File

@@ -0,0 +1,123 @@
from joysim.annotations.config_class import configclass, field
from joysim.annotations.stereotype import stereotype
from joysim.app import JoySim
from joysim.core.scene_manager import SceneManager
from joysim.extensions.benchmark.action import RobotAction
from joysim.extensions.benchmark.benchmark import (
BenchmarkAction,
BenchmarkObservation,
ControlMode,
)
from joysim.extensions.benchmark.policy import Policy, PolicyConfig
import numpy as np
import pickle
import requests
@configclass
@stereotype.register_config("starvla")
class StarvlaPolicyConfig(PolicyConfig):
robot_name: str = field(default="my_robot", required=True, comment="The name of the robot")
object_name: str = field(default="target", required=True, comment="The name of the object")
server_url: str = field(
default="http://127.0.0.1:5000/policy",
required=True,
comment="StarVLA policy server url"
)
prompt: str = field(
default="pick the object",
required=True,
comment="task instruction"
)
@stereotype.register_model("starvla")
class StarvlaPolicy(Policy):
def __init__(self, config: StarvlaPolicyConfig):
super().__init__(config)
self.robot_name = config.robot_name
self.object_name = config.object_name
self.server_url = config.server_url
self.prompt = config.prompt
def reset(self) -> None:
pass
def warmup(self, benchmark_observation: BenchmarkObservation) -> None:
pass
def needs_observation(self) -> bool:
return True
def preprocess_observation(self, benchmark_observation: BenchmarkObservation) -> dict:
robot_obs = benchmark_observation.get_robot_observations(self.robot_name)["robot_data"]
joint_positions = robot_obs["joint_positions"]
robot_position = robot_obs["position"]
robot_quaternion = robot_obs["rotation"]
state = np.concatenate([
robot_position,
robot_quaternion,
np.array([0.0])
])
camera_obs = benchmark_observation.get_sensor_observations()
rgb = camera_obs["rgb"]
obs = {
"state": np.expand_dims(state, axis=0),
"joint": np.expand_dims(joint_positions, axis=0),
"rgb": np.expand_dims(rgb, axis=0),
"prompt": self.prompt
}
return obs
def compute_action(self, observation: dict) -> dict:
payload = pickle.dumps(observation)
response = requests.post(
self.server_url,
data=payload,
headers={"Content-Type": "application/octet-stream"}
)
if response.status_code != 200:
raise RuntimeError(f"StarVLA server error: {response.text}")
result = pickle.loads(response.content)
return result
def postprocess_action(self, action: dict) -> BenchmarkAction:
benchmark_action = BenchmarkAction()
robot = SceneManager.get_robot(self.robot_name)
joint_names = robot.get_planner().get_plannable_joint_names()
joint_positions = action["action"][0]
benchmark_action.add_robot_action(
RobotAction(
control_mode=ControlMode.POSITION,
robot_name=self.robot_name,
joint_names=joint_names,
joint_positions=joint_positions
)
)
return benchmark_action
if __name__ == "__main__":
js = JoySim("./benchmark.yaml")
js.start()