File size: 4,217 Bytes
bf63c73 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | """Process-local wrist-camera retrofit for every 2/3/4-agent RoboFactory task.
RoboFactory's YAML files declare world-fixed cameras, even when their uid says
``head_camera_agent{i}``. This module leaves those files untouched and changes
only the live CameraConfig objects: local camera i is mounted on Panda i's
``panda_hand`` link. Collection, training rollout, and evaluation must import
this same module so the observation distribution is identical end-to-end.
"""
from __future__ import annotations
import os
import numpy as np
import sapien
# This is the simulated equivalent of adding a ``camera_link`` to the Panda
# URDF with a fixed joint whose parent is ``panda_hand``. It is intentionally
# identical for every Panda in every task: the lower edge keeps the gripper
# fingertips in view while the camera follows that hand rigidly.
CAMERA_LINK_IN_HAND = sapien.Pose(
p=[0.0465, -0.0200, 0.0360], q=[0.0, 0.70710678, 0.0, 0.70710678]
)
# Backwards-compatible alias used by preview utilities.
LOCAL_POSE = CAMERA_LINK_IN_HAND
# Match RoboFactory's native 320x240, 90-degree pinhole camera. A prior
# 135-degree experiment made the peripheral perspective visibly stretched.
CAMERA_FOV = 1.5707963268
# The conventional ACT corpus stays at RoboFactory's 320×240 default. The
# future frozen-DINO experiment sets these *before importing this module* to
# archive genuine 640×480 wrist frames instead of upsampling old RGB data.
CAMERA_WIDTH = int(os.environ.get("ROBOFACTORY_WRIST_WIDTH", "320"))
CAMERA_HEIGHT = int(os.environ.get("ROBOFACTORY_WRIST_HEIGHT", "240"))
CAMERA_NEAR = 0.01
CAMERA_FAR = 10.0
# PlaceFood starts with both wrists on the far side of a large pot. Turning
# the physical hand camera toward the interior side of the wrist exposes the
# pot/meat workspace without making the camera world-fixed.
def _install(task_class, num_agents: int):
stock_property = task_class._default_sensor_configs
def wrist_sensor_configs(self):
# A decentralized policy must never receive or archive a global camera
# stream. Human-render cameras remain separate from these observations.
configs = [config for config in stock_property.fget(self)
if config.uid != "head_camera_global"]
local = {f"head_camera_agent{index}": index for index in range(num_agents)}
for config in configs:
agent_index = local.get(config.uid)
if agent_index is None:
continue
hand = next(
link for link in self.agent.agents[agent_index].robot.links
if link.name == "panda_hand"
)
config.mount = hand
config.pose = CAMERA_LINK_IN_HAND
# All policy cameras share exactly one pinhole projection model.
config.width = CAMERA_WIDTH
config.height = CAMERA_HEIGHT
config.fov = CAMERA_FOV
config.near = CAMERA_NEAR
config.far = CAMERA_FAR
return configs
task_class._default_sensor_configs = property(wrist_sensor_configs)
def install_all():
"""Install once per Python process; idempotent for normal script imports."""
from robofactory.tasks.camera_alignment import CameraAlignmentEnv
from robofactory.tasks.lift_barrier import LiftBarrierEnv
from robofactory.tasks.pass_shoe import PassShoeEnv
from robofactory.tasks.place_food import PlaceFoodEnv
from robofactory.tasks.three_robots_stack_cube import ThreeRobotsStackCubeEnv
from robofactory.tasks.two_robots_stack_cube import TwoRobotsStackCubeEnv
from robofactory.tasks.long_pipeline_delivery import LongPipelineDeliveryEnv
from robofactory.tasks.take_photo import TakePhotoEnv
for task_class, count in (
(LiftBarrierEnv, 2),
(PassShoeEnv, 2),
(PlaceFoodEnv, 2),
(TwoRobotsStackCubeEnv, 2),
(CameraAlignmentEnv, 3),
(ThreeRobotsStackCubeEnv, 3),
(LongPipelineDeliveryEnv, 4),
(TakePhotoEnv, 4),
):
if not getattr(task_class, "_wrist_camera_patch_installed", False):
_install(task_class, count)
task_class._wrist_camera_patch_installed = True
install_all()
|