Fix XR teleop: body-frame IK control for mobile chassis

Switch arm IK from world-frame to body-frame control so that
pushing the XR controller forward always moves the arm in the
robot's forward direction, regardless of chassis rotation.

Key changes:
- dual_arm_agent: convert EEF observations to body frame before
  passing to XR controller; send body-frame IK targets directly
  (removed convert_action_world_to_root)
- xr_controller: XR deltas treated as body-frame deltas (no yaw
  rotation needed — VR view tracks robot heading naturally)
- streaming: add debug frame save for stereo alignment diagnostics
- mindrobot_2i_cfg: IdealPDActuator for trunk, disabled gravity
- Author headers updated to Yutang Li, SIAT

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-03-25 11:32:28 +08:00
parent 05a146bca6
commit bfe28b188a
12 changed files with 82 additions and 882 deletions

View File

@@ -1,856 +0,0 @@
#!/usr/bin/env python3
# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
# All rights reserved.
#
# SPDX-License-Identifier: BSD-3-Clause
"""
Script to run teleoperation with Isaac Lab manipulation environments using PICO XR Controllers.
Only absolute IK mode is supported (Isaac-MindRobot-LeftArm-IK-Abs-v0).
The controller computes delta pose in Isaac Sim world frame and accumulates
an absolute EEF target, which is then converted to robot root frame before
being sent to the DifferentialIK action manager.
"""
import argparse
import logging
import sys
import os
from collections.abc import Callable
# Ensure xr_utils (next to this script) is importable when running directly
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from isaaclab.app import AppLauncher
logger = logging.getLogger(__name__)
# add argparse arguments
parser = argparse.ArgumentParser(
description="Teleoperation for Isaac Lab environments with PICO XR Controller (absolute IK mode only)."
)
parser.add_argument("--num_envs", type=int, default=1, help="Number of environments to simulate.")
parser.add_argument(
"--task",
type=str,
default="Isaac-MindRobot-LeftArm-IK-Abs-v0",
help="Name of the task (must be an Abs IK task).",
)
parser.add_argument(
"--sensitivity", type=float, default=None,
help="Set both pos and rot sensitivity (overridden by --pos_sensitivity/--rot_sensitivity).",
)
parser.add_argument("--pos_sensitivity", type=float, default=None, help="Position sensitivity. Default: 1.0.")
parser.add_argument("--rot_sensitivity", type=float, default=None, help="Rotation sensitivity. Default: 0.3.")
parser.add_argument("--arm", type=str, default="left", choices=["left", "right"], help="Which arm/controller to use.")
parser.add_argument("--base_speed", type=float, default=5.0, help="Max wheel speed (rad/s) for joystick full forward.")
parser.add_argument("--base_turn", type=float, default=2.0, help="Max wheel differential (rad/s) for joystick full left/right.")
parser.add_argument("--drive_speed", type=float, default=0.5, help="Max robot linear speed (m/s) for direct base control. Default: 0.5")
parser.add_argument("--drive_turn", type=float, default=1.5, help="Max robot yaw rate (rad/s) for direct base control. Default: 1.5")
parser.add_argument(
"--stream-to", type=str, default=None, dest="stream_to",
help="PICO VR headset IP for sim stereo streaming. Streams H264 SBS to the Unity client TCP server. "
"Requires vr_left_eye/vr_right_eye cameras in the env scene config.",
)
parser.add_argument("--stream-port", type=int, default=12345, dest="stream_port", help="TCP port for video streaming.")
parser.add_argument("--stream-bitrate", type=int, default=20_000_000, dest="stream_bitrate", help="H264 bitrate (bps).")
parser.add_argument(
"--debug-viewports", action="store_true", dest="debug_viewports", default=False,
help="Enable debug viewport windows (Left Hand, Right Hand, Head, Chest). Off by default when --stream-to is used.",
)
AppLauncher.add_app_launcher_args(parser)
args_cli = parser.parse_args()
app_launcher_args = vars(args_cli)
app_launcher_args["xr"] = False
app_launcher = AppLauncher(app_launcher_args)
simulation_app = app_launcher.app
"""Rest everything follows."""
import gymnasium as gym
import numpy as np
import torch
from scipy.spatial.transform import Rotation as R
from isaaclab.envs import ManagerBasedRLEnvCfg
import isaaclab_tasks # noqa: F401
import mindbot.tasks # noqa: F401
from isaaclab_tasks.utils import parse_env_cfg
from xr_utils import XrClient, transform_xr_pose, quat_diff_as_rotvec_xyzw, is_valid_quaternion
from xr_utils.geometry import R_HEADSET_TO_WORLD
# =====================================================================
# Robot Camera Viewport Utilities
# =====================================================================
# Resolved prim paths for env_0 — must match CameraCfg prim_path with
# {ENV_REGEX_NS} → /World/envs/env_0
_ROBOT_CAM_PRIMS: dict[str, str] = {
"Left Hand": "/World/envs/env_0/Robot/rm_65_fb_left/Link_6/robot___left_8_02/cam_left_hand",
"Right Hand": "/World/envs/env_0/Robot/rm_65_b_right/Link6/robot___right_8_2/cam_right_hand",
"Head": "/World/envs/env_0/Robot/robot_head/ZED_X/base_link/ZED_X/CameraLeft",
"Chest": "/World/envs/env_0/Robot/robot_trunk/cam_chest",
}
# Stores created viewport window objects
_robot_viewports: dict[str, object] = {}
def create_robot_viewports() -> None:
"""Create 4 viewport windows and bind each to a robot camera prim.
Must be called after env.reset() so all prims exist on the USD stage.
"""
try:
import omni.kit.viewport.utility as vp_util
except ImportError:
logger.warning("[Viewport] omni.kit.viewport.utility not available — skipping viewport creation.")
return
for name, cam_path in _ROBOT_CAM_PRIMS.items():
vp_win = vp_util.create_viewport_window(
f"Robot {name} View", width=640, height=360
)
vp_win.viewport_api.camera_path = cam_path
_robot_viewports[name] = vp_win
print(f"[INFO] Viewport 'Robot {name} View' bound to: {cam_path}")
def rebind_robot_viewports() -> None:
"""Re-bind all robot viewports to their camera paths.
Call this after every env reset so viewports stay locked to the cameras.
"""
for name, vp_win in _robot_viewports.items():
cam_path = _ROBOT_CAM_PRIMS[name]
vp_win.viewport_api.camera_path = cam_path
# =====================================================================
# Helpers
# =====================================================================
def _quat_wxyz_to_rotation(quat_wxyz: np.ndarray) -> R:
"""Convert Isaac-style wxyz quaternion to scipy Rotation."""
return R.from_quat([quat_wxyz[1], quat_wxyz[2], quat_wxyz[3], quat_wxyz[0]])
def _rotation_to_quat_wxyz(rot: R) -> np.ndarray:
"""Convert scipy Rotation to Isaac-style wxyz quaternion."""
q = rot.as_quat()
return np.array([q[3], q[0], q[1], q[2]])
def world_pose_to_root_frame(
pos_w: np.ndarray,
quat_wxyz: np.ndarray,
root_pos_w: np.ndarray,
root_quat_wxyz: np.ndarray,
) -> tuple[np.ndarray, np.ndarray]:
"""Express a world-frame pose in the robot root frame."""
root_rot = _quat_wxyz_to_rotation(root_quat_wxyz)
pose_rot = _quat_wxyz_to_rotation(quat_wxyz)
pos_root = root_rot.inv().apply(pos_w - root_pos_w)
quat_root = _rotation_to_quat_wxyz(root_rot.inv() * pose_rot)
return pos_root, quat_root
# =====================================================================
# Teleoperation Controller
# =====================================================================
class XrTeleopController:
"""Teleop controller for PICO XR headset (absolute IK mode).
Accumulates an absolute EEF target in Isaac Sim world frame.
Grip button acts as clutch; B/Y buttons trigger environment reset.
"""
def __init__(self, arm="left", pos_sensitivity=1.0, rot_sensitivity=0.3,
max_pos_per_step=0.05, max_rot_per_step=0.08, xr_client=None):
self.xr_client = xr_client if xr_client is not None else XrClient()
self._owns_client = xr_client is None # only close if we created it
self.pos_sensitivity = pos_sensitivity
self.rot_sensitivity = rot_sensitivity
# Hard caps per physics step to prevent IK divergence.
# max_rot_per_step ~4.6°, max_pos_per_step 5 cm.
self.max_pos_per_step = max_pos_per_step
self.max_rot_per_step = max_rot_per_step
self.arm = arm
self.controller_name = "left_controller" if arm == "left" else "right_controller"
self.grip_name = "left_grip" if arm == "left" else "right_grip"
self.trigger_name = "left_trigger" if arm == "left" else "right_trigger"
self.R_headset_world = R_HEADSET_TO_WORLD
self.prev_xr_pos = None
self.prev_xr_quat = None
self.target_eef_pos = None # world frame
self.target_eef_quat = None # world frame, wxyz
self.grip_active = False
self.frame_count = 0
self.reset_button_latched = False
self.require_grip_reengage = False
self.grip_engage_threshold = 0.8
self.grip_release_threshold = 0.2
self.callbacks = {}
def add_callback(self, name: str, func: Callable):
self.callbacks[name] = func
def reset(self):
self.prev_xr_pos = None
self.prev_xr_quat = None
self.grip_active = False
self.frame_count = 0
self.target_eef_pos = None
self.target_eef_quat = None
# Require grip release after reset to avoid driving to stale pose.
self.require_grip_reengage = True
def close(self):
if self._owns_client:
self.xr_client.close()
def get_zero_action(self, trigger, current_eef_pos=None, current_eef_quat=None) -> torch.Tensor:
"""Return 8D hold action [x, y, z, qw, qx, qy, qz, gripper] at frozen target pose."""
action = torch.zeros(8)
if self.target_eef_pos is not None and self.target_eef_quat is not None:
# Prefer frozen world-frame target so IK holds a fixed world position
# even if the robot chassis drifts slightly under gravity.
action[:3] = torch.tensor(self.target_eef_pos.copy())
action[3:7] = torch.tensor(self.target_eef_quat.copy())
elif current_eef_pos is not None and current_eef_quat is not None:
action[:3] = torch.tensor(current_eef_pos.copy())
action[3:7] = torch.tensor(current_eef_quat.copy())
else:
action[3] = 1.0 # identity quaternion
action[7] = 1.0 if trigger > 0.5 else -1.0
return action
def advance(self, current_eef_pos=None, current_eef_quat=None) -> torch.Tensor:
"""Read XR controller and return 8D absolute action [x, y, z, qw, qx, qy, qz, gripper].
Position and quaternion are in Isaac Sim world frame.
Caller must convert to robot root frame before sending to IK.
"""
try:
# Only left controller Y button resets (not right controller B)
if self.arm == "left":
reset_pressed = self.xr_client.get_button("Y")
else:
reset_pressed = False
if reset_pressed and not self.reset_button_latched:
if "RESET" in self.callbacks:
self.callbacks["RESET"]()
self.reset_button_latched = reset_pressed
except Exception:
pass
try:
raw_pose = self.xr_client.get_pose(self.controller_name)
grip = self.xr_client.get_key_value(self.grip_name)
trigger = self.xr_client.get_key_value(self.trigger_name)
except Exception:
return self.get_zero_action(0.0, current_eef_pos, current_eef_quat)
if not is_valid_quaternion(raw_pose[3:]):
return self.get_zero_action(trigger, current_eef_pos, current_eef_quat)
# Wait for grip release after reset
if self.require_grip_reengage:
if grip <= self.grip_release_threshold:
self.require_grip_reengage = False
else:
if current_eef_pos is not None and current_eef_quat is not None:
self.target_eef_pos = current_eef_pos.copy()
self.target_eef_quat = current_eef_quat.copy()
return self.get_zero_action(trigger, current_eef_pos, current_eef_quat)
# Grip button as clutch with hysteresis
grip_pressed = grip > self.grip_release_threshold if self.grip_active else grip >= self.grip_engage_threshold
if not grip_pressed:
self.prev_xr_pos = None
self.prev_xr_quat = None
# Only snapshot target on the transition frame (grip just released) or first ever frame.
# After that, keep it frozen so IK holds a fixed world-frame position.
if self.grip_active or self.target_eef_pos is None:
if current_eef_pos is not None and current_eef_quat is not None:
self.target_eef_pos = current_eef_pos.copy()
self.target_eef_quat = current_eef_quat.copy()
self.grip_active = False
return self.get_zero_action(trigger, current_eef_pos, current_eef_quat)
if not self.grip_active:
self.prev_xr_pos = raw_pose[:3].copy()
self.prev_xr_quat = raw_pose[3:].copy()
if current_eef_pos is not None and current_eef_quat is not None:
self.target_eef_pos = current_eef_pos.copy()
self.target_eef_quat = current_eef_quat.copy()
self.grip_active = True
return self.get_zero_action(trigger, current_eef_pos, current_eef_quat)
# Project current and previous XR poses into Isaac Sim world frame
xr_world_pos, xr_world_quat_xyzw = transform_xr_pose(raw_pose[:3], raw_pose[3:])
prev_xr_world_pos, prev_xr_world_quat_xyzw = transform_xr_pose(self.prev_xr_pos, self.prev_xr_quat)
# Delta position (world frame), clamped per step
world_delta_pos = (xr_world_pos - prev_xr_world_pos) * self.pos_sensitivity
pos_norm = np.linalg.norm(world_delta_pos)
if pos_norm > self.max_pos_per_step:
world_delta_pos *= self.max_pos_per_step / pos_norm
# Delta rotation (world frame), clamped per step
world_delta_rot = quat_diff_as_rotvec_xyzw(prev_xr_world_quat_xyzw, xr_world_quat_xyzw) * self.rot_sensitivity
rot_norm = np.linalg.norm(world_delta_rot)
if rot_norm > self.max_rot_per_step:
world_delta_rot *= self.max_rot_per_step / rot_norm
gripper_action = 1.0 if trigger > 0.5 else -1.0
# Accumulate absolute target
if self.target_eef_pos is None:
self.target_eef_pos = np.zeros(3)
self.target_eef_quat = np.array([1.0, 0.0, 0.0, 0.0])
self.target_eef_pos += world_delta_pos
# Apply rotation delta in world frame so controller direction = EEF world direction
target_r = _quat_wxyz_to_rotation(self.target_eef_quat)
self.target_eef_quat = _rotation_to_quat_wxyz(R.from_rotvec(world_delta_rot) * target_r)
self.prev_xr_pos = raw_pose[:3].copy()
self.prev_xr_quat = raw_pose[3:].copy()
self.frame_count += 1
action = torch.tensor([
*self.target_eef_pos,
*self.target_eef_quat,
gripper_action,
], dtype=torch.float32)
if self.frame_count % 30 == 0:
np.set_printoptions(precision=4, suppress=True, floatmode='fixed')
print("\n====================== [VR TELEOP DEBUG] ======================")
print(f"| Raw VR Pos (OpenXR): {np.array(raw_pose[:3])}")
print(f"| Raw VR Quat (xyzw): {np.array(raw_pose[3:])}")
print(f"| XR Delta Pos (world): {world_delta_pos} (norm={np.linalg.norm(world_delta_pos):.4f})")
print(f"| XR Delta Rot (world): {world_delta_rot} (norm={np.linalg.norm(world_delta_rot):.4f})")
print(f"| Targ Pos (world): {action[:3].numpy()}")
print(f"| Targ Quat (world, wxyz): {action[3:7].numpy()}")
print(f"| Gripper & Trigger: Grip={grip:.2f}, Trig={trigger:.2f}")
print("===============================================================")
return action
# =====================================================================
# Head Tracking Utility
# =====================================================================
def get_head_joint_targets(
xr_client, neutral_rot: R | None
) -> tuple[np.ndarray, R | None]:
"""Extract head joint targets [yaw, pitch] from HMD pose.
At first call (neutral_rot is None) the current headset orientation
is recorded as the neutral reference and zeros are returned.
Subsequent calls return yaw/pitch relative to that neutral pose.
Returns:
(joint_targets, neutral_rot) where joint_targets is shape (2,)
float32 [head_yaw_Joint_rad, head_pitch_Joint_rad].
"""
try:
raw_pose = xr_client.get_pose("headset")
if not is_valid_quaternion(raw_pose[3:]):
return np.zeros(2, dtype=np.float32), neutral_rot
_, head_world_quat_xyzw = transform_xr_pose(raw_pose[:3], raw_pose[3:])
head_rot = R.from_quat(head_world_quat_xyzw)
if neutral_rot is None:
# First call — record neutral and hold zero
return np.zeros(2, dtype=np.float32), head_rot
# Relative rotation from neutral heading
rel_rot = head_rot * neutral_rot.inv()
# ZYX Euler: [0]=yaw (Z), [1]=pitch (Y)
euler_zyx = rel_rot.as_euler("ZYX")
yaw = float(np.clip(euler_zyx[0], -1.57, 1.57)) # ±90°
pitch = float(np.clip(euler_zyx[1], -0.52, 0.52)) # ±30°
return np.array([yaw, pitch], dtype=np.float32), neutral_rot
except Exception:
return np.zeros(2, dtype=np.float32), neutral_rot
# =====================================================================
# Main Execution Loop
# =====================================================================
def main() -> None:
"""Run teleoperation with PICO XR Controller against Isaac Lab environment."""
env_cfg = parse_env_cfg(args_cli.task, num_envs=args_cli.num_envs)
env_cfg.env_name = args_cli.task
if not isinstance(env_cfg, ManagerBasedRLEnvCfg):
raise ValueError(f"Teleoperation requires ManagerBasedRLEnvCfg. Got: {type(env_cfg)}")
if "Abs" not in args_cli.task:
raise ValueError(
f"Task '{args_cli.task}' is not an absolute IK task. "
"Only Abs tasks are supported (e.g. Isaac-MindRobot-LeftArm-IK-Abs-v0)."
)
env_cfg.terminations.time_out = None
try:
env = gym.make(args_cli.task, cfg=env_cfg).unwrapped
except Exception as e:
logger.error(f"Failed to create environment '{args_cli.task}': {e}")
simulation_app.close()
return
# Detect single-arm vs dual-arm based on action manager term names
is_dual_arm = "left_arm_action" in env.action_manager._terms
has_head = "head_action" in env.action_manager._terms
# -- Sim-to-VR stereo streaming setup --
streamer = None
if args_cli.stream_to:
try:
from mindbot.utils.sim_video_streamer import SimVideoStreamer
# Get camera resolution from scene config (vr_left_eye)
scene = env.unwrapped.scene
if "vr_left_eye" not in scene.sensors:
logger.warning("[Stream] vr_left_eye camera not found in scene. Streaming disabled.")
else:
cam = scene["vr_left_eye"]
cam_w = cam.cfg.width
cam_h = cam.cfg.height
streamer = SimVideoStreamer(
host=args_cli.stream_to,
port=args_cli.stream_port,
width=cam_w,
height=cam_h,
fps=30,
bitrate=args_cli.stream_bitrate,
)
if streamer.connect():
print(f"[INFO] Sim stereo streaming to {args_cli.stream_to}:{args_cli.stream_port} "
f"({cam_w*2}x{cam_h} SBS @ 30fps)")
else:
logger.error("[Stream] Failed to connect. Streaming disabled.")
streamer = None
except ImportError:
logger.error("[Stream] mindbot.utils.sim_video_streamer not found. pip install av?")
except Exception as e:
logger.error(f"[Stream] Init failed: {e}")
# Sensitivity: explicit args override --sensitivity, which overrides defaults
pos_sens = 1.0
rot_sens = 0.3
if args_cli.sensitivity is not None:
pos_sens = args_cli.sensitivity
rot_sens = args_cli.sensitivity
if args_cli.pos_sensitivity is not None:
pos_sens = args_cli.pos_sensitivity
if args_cli.rot_sensitivity is not None:
rot_sens = args_cli.rot_sensitivity
if is_dual_arm:
print(f"\n[INFO] Dual-arm mode detected. Using both controllers.")
print(f"[INFO] IK Mode: ABSOLUTE")
print(f"[INFO] Sensitivity: pos={pos_sens:.3f} rot={rot_sens:.3f}")
shared_client = XrClient()
teleop_left = XrTeleopController(arm="left", pos_sensitivity=pos_sens, rot_sensitivity=rot_sens, xr_client=shared_client)
teleop_right = XrTeleopController(arm="right", pos_sensitivity=pos_sens, rot_sensitivity=rot_sens, xr_client=shared_client)
teleop_interface = teleop_left # primary interface for callbacks/reset
teleop_right_ref = teleop_right
else:
print(f"\n[INFO] Connecting to PICO XR Headset using {args_cli.arm} controller...")
print(f"[INFO] IK Mode: ABSOLUTE")
print(f"[INFO] Sensitivity: pos={pos_sens:.3f} rot={rot_sens:.3f}")
teleop_interface = XrTeleopController(
arm=args_cli.arm,
pos_sensitivity=pos_sens,
rot_sensitivity=rot_sens,
)
should_reset = False
def request_reset():
nonlocal should_reset
should_reset = True
print("[INFO] Reset requested via XR button.")
teleop_interface.add_callback("RESET", request_reset)
def clear_ik_target_state():
"""Sync IK controller's internal target to current EEF pose to avoid jumps after reset."""
if is_dual_arm:
for term_key in ["left_arm_action", "right_arm_action"]:
arm_action_term = env.action_manager._terms[term_key]
ee_pos_b, ee_quat_b = arm_action_term._compute_frame_pose()
arm_action_term._raw_actions.zero_()
arm_action_term._processed_actions.zero_()
arm_action_term._ik_controller._command.zero_()
arm_action_term._ik_controller.ee_pos_des[:] = ee_pos_b
arm_action_term._ik_controller.ee_quat_des[:] = ee_quat_b
else:
arm_action_term = env.action_manager._terms["arm_action"]
ee_pos_b, ee_quat_b = arm_action_term._compute_frame_pose()
arm_action_term._raw_actions.zero_()
arm_action_term._processed_actions.zero_()
arm_action_term._ik_controller._command.zero_()
arm_action_term._ik_controller.ee_pos_des[:] = ee_pos_b
arm_action_term._ik_controller.ee_quat_des[:] = ee_quat_b
def convert_action_world_to_root(action_tensor: torch.Tensor) -> torch.Tensor:
"""Convert absolute IK action from world frame to robot root frame."""
robot = env.unwrapped.scene["robot"]
root_pos_w = robot.data.root_pos_w[0].detach().cpu().numpy()
root_quat_w = robot.data.root_quat_w[0].detach().cpu().numpy()
pos_root, quat_root = world_pose_to_root_frame(
action_tensor[:3].numpy(), action_tensor[3:7].numpy(),
root_pos_w, root_quat_w,
)
out = action_tensor.clone()
out[:3] = torch.tensor(pos_root, dtype=torch.float32)
out[3:7] = torch.tensor(quat_root, dtype=torch.float32)
return out
def get_chassis_commands() -> tuple[torch.Tensor, float, float]:
"""Read left joystick → (wheel_cmd 4D, v_fwd m/s, omega rad/s).
wheel_cmd : differential wheel velocities for the action vector (training data).
v_fwd : desired robot forward speed in m/s (for direct base control).
omega : desired robot yaw rate in rad/s (for direct base control).
"""
try:
joy = teleop_interface.xr_client.get_joystick("left")
jy = float(joy[1])
jx = float(joy[0])
except Exception:
return torch.zeros(4), 0.0, 0.0
# Wheel velocity commands for action vector (rad/s)
v_w = jy * args_cli.base_speed
omega_w = jx * args_cli.base_turn
right_vel = v_w - omega_w
left_vel = v_w + omega_w
wheel_cmd = torch.tensor([right_vel, left_vel, left_vel, right_vel], dtype=torch.float32)
if abs(v_w) > 0.1 or abs(omega_w) > 0.1:
print(f"[WHEEL] joy=({jx:+.2f},{jy:+.2f}) v_wheel={v_w:.2f} ω_wheel={omega_w:.2f}{wheel_cmd.tolist()}")
# Direct base velocity commands (m/s, rad/s)
v_fwd = jy * args_cli.drive_speed
omega = -jx * args_cli.drive_turn # jx<0 (left push) → positive yaw = left turn
return wheel_cmd, v_fwd, omega
def apply_base_velocity(v_fwd: float, omega: float) -> None:
"""Directly set robot root linear+angular velocity to bypass isotropic friction.
PhysX uses isotropic friction, so skid-steer lateral slip is impossible through
wheel torques alone. This function overrides the base velocity each step to give
clean translational + rotational motion regardless of friction.
The write is buffered and applied at the start of the next env.step().
"""
if abs(v_fwd) < 1e-4 and abs(omega) < 1e-4:
return
robot = env.unwrapped.scene["robot"]
rq = robot.data.root_quat_w # [N, 4] wxyz
w_q, x_q, y_q, z_q = rq[:, 0], rq[:, 1], rq[:, 2], rq[:, 3]
# Robot forward direction in world frame (rotate [1,0,0] by root quaternion)
fwd_x = 1.0 - 2.0 * (y_q * y_q + z_q * z_q)
fwd_y = 2.0 * (x_q * y_q + w_q * z_q)
vel = torch.zeros(env.num_envs, 6, device=device)
vel[:, 0] = v_fwd * fwd_x # world x linear
vel[:, 1] = v_fwd * fwd_y # world y linear
# vel[:, 2] = 0 # z: let physics handle gravity/contact
vel[:, 5] = omega # world z angular (yaw)
robot.write_root_velocity_to_sim(vel)
obs, _ = env.reset()
clear_ik_target_state()
teleop_interface.reset()
# Read contact sensor data directly from OmniGraph node USD attributes.
# The USD has action graphs that execute each physics step and write outputs to:
# outputs:inContact (bool) outputs:value (float, Newtons)
# /root in USD → /World/envs/env_0/Robot in the loaded scene.
_CONTACT_OG_NODES = {
"left_r": (
"/World/envs/env_0/Robot"
"/l_zhuajia___m2_1_cs/isaac_read_contact_sensor_node"
),
"left_l": (
"/World/envs/env_0/Robot"
"/l_zhuajia___m3_1_cs/isaac_read_contact_sensor_node"
),
}
def _og_contact_reading(stage, node_path: str) -> tuple[bool, float]:
"""Read one OmniGraph contact node output via USD attribute API."""
prim = stage.GetPrimAtPath(node_path)
if not prim.IsValid():
return False, float("nan")
in_contact_attr = prim.GetAttribute("outputs:inContact")
force_attr = prim.GetAttribute("outputs:value")
in_contact = in_contact_attr.Get() if in_contact_attr.IsValid() else False
force = force_attr.Get() if force_attr.IsValid() else 0.0
return bool(in_contact), float(force or 0.0)
# One-time diagnostic: run after first env.reset() so all prims exist
def _diag_contact_nodes() -> None:
import omni.usd
stage = omni.usd.get_context().get_stage()
print("\n[DIAG] OmniGraph contact node check:")
for name, path in _CONTACT_OG_NODES.items():
prim = stage.GetPrimAtPath(path)
if prim.IsValid():
has_in = prim.GetAttribute("outputs:inContact").IsValid()
has_val = prim.GetAttribute("outputs:value").IsValid()
print(f" [{name}] ✓ prim valid | outputs:inContact={has_in} outputs:value={has_val}")
else:
# Try to find similar prims to hint correct path
parent_path = "/".join(path.split("/")[:-1])
parent = stage.GetPrimAtPath(parent_path)
print(f" [{name}] ✗ NOT FOUND at: {path}")
if parent.IsValid():
children = [c.GetName() for c in parent.GetChildren()]
print(f" parent exists, children: {children}")
else:
gp_path = "/".join(path.split("/")[:-2])
gp = stage.GetPrimAtPath(gp_path)
print(f" grandparent '{gp_path}' valid={gp.IsValid()}")
def read_gripper_contact() -> str:
"""Return contact status string by reading OmniGraph node USD attributes."""
try:
import omni.usd
stage = omni.usd.get_context().get_stage()
c_r, f_r = _og_contact_reading(stage, _CONTACT_OG_NODES["left_r"])
c_l, f_l = _og_contact_reading(stage, _CONTACT_OG_NODES["left_l"])
if f_r != f_r: # nan → prim not found
return "prim not found (run with diag)"
sym_r = "" if c_r else ""
sym_l = "" if c_l else ""
return f"{sym_r}R={f_r:.1f}N {sym_l}L={f_l:.1f}N"
except Exception as _e:
return f"N/A ({_e})"
# One-time diagnostic: verify OmniGraph contact sensor node paths
_diag_contact_nodes()
# Create debug viewport windows (always on)
create_robot_viewports()
if is_dual_arm:
teleop_right_ref.reset()
print("\n" + "=" * 50)
print(" Teleoperation Started!")
if is_dual_arm:
print(" LEFT controller → left arm")
print(" RIGHT controller → right arm")
print(" TRIGGER: open/close gripper")
print(" GRIP (hold): move the arm")
print(" Left joystick: Y=forward/back, X=turn")
print(" Y (left controller): reset environment")
print("=" * 50 + "\n")
device = env.unwrapped.device
sim_frame = 0
obs, _ = env.reset()
clear_ik_target_state()
last_root_left = None
last_root_right = None
neutral_head_rot = None # calibrated on first headset sample after reset
# trunk_action uses use_default_offset=False, so action = absolute position target.
trunk_cmd = torch.tensor([0.1], dtype=torch.float32)
while simulation_app.is_running():
try:
with torch.inference_mode():
if should_reset:
obs, _ = env.reset()
clear_ik_target_state()
teleop_interface.reset()
if is_dual_arm:
teleop_right_ref.reset()
rebind_robot_viewports()
should_reset = False
sim_frame = 0
last_root_left = None
last_root_right = None
neutral_head_rot = None # re-calibrate on next headset sample
continue
policy_obs = obs["policy"]
wheel_np, v_fwd, omega_base = get_chassis_commands()
# Head tracking: HMD yaw/pitch relative to neutral pose
head_cmd_np, neutral_head_rot = get_head_joint_targets(
teleop_interface.xr_client, neutral_head_rot
)
head_cmd = torch.tensor(head_cmd_np, dtype=torch.float32)
if is_dual_arm:
eef_pos_left = policy_obs["eef_pos_left"][0].cpu().numpy()
eef_quat_left = policy_obs["eef_quat_left"][0].cpu().numpy()
eef_pos_right = policy_obs["eef_pos_right"][0].cpu().numpy()
eef_quat_right = policy_obs["eef_quat_right"][0].cpu().numpy()
left_action = teleop_left.advance(current_eef_pos=eef_pos_left, current_eef_quat=eef_quat_left)
right_action = teleop_right_ref.advance(current_eef_pos=eef_pos_right, current_eef_quat=eef_quat_right)
# Only recompute root-frame IK target when grip is active; otherwise freeze joints
if teleop_left.grip_active or last_root_left is None:
last_root_left = convert_action_world_to_root(left_action)[:7].clone()
if teleop_right_ref.grip_active or last_root_right is None:
last_root_right = convert_action_world_to_root(right_action)[:7].clone()
# Action layout depends on task:
# with head: left_arm(7) | wheel(4) | left_gripper(1) | right_arm(7) | right_gripper(1) | head(2) | trunk(1) = 23D
# without head: left_arm(7) | wheel(4) | left_gripper(1) | right_arm(7) | right_gripper(1) = 20D
parts = [
last_root_left, wheel_np, left_action[7:8],
last_root_right, right_action[7:8],
]
if has_head:
parts.append(head_cmd)
parts.append(trunk_cmd)
action_np = torch.cat(parts)
else:
eef_pos = policy_obs["eef_pos"][0].cpu().numpy()
eef_quat = policy_obs["eef_quat"][0].cpu().numpy()
action_np = teleop_interface.advance(current_eef_pos=eef_pos, current_eef_quat=eef_quat)
if teleop_interface.grip_active or last_root_left is None:
last_root_left = convert_action_world_to_root(action_np)[:7].clone()
# Action layout: arm(7) | wheel(4) | gripper(1)
action_np = torch.cat([last_root_left, wheel_np, action_np[7:8]])
actions = action_np.unsqueeze(0).repeat(env.num_envs, 1).to(device)
obs, _, _, _, _ = env.step(actions)
# Direct base velocity override: bypasses isotropic friction limitation
# so skid-steer turning works correctly.
apply_base_velocity(v_fwd, omega_base)
# Stream sim stereo to VR headset (if enabled)
if streamer is not None:
try:
scene = env.unwrapped.scene
left_rgb = scene["vr_left_eye"].data.output["rgb"][0].cpu().numpy()
right_rgb = scene["vr_right_eye"].data.output["rgb"][0].cpu().numpy()
if not streamer.send_frame(left_rgb, right_rgb):
logger.warning("[Stream] Connection lost. Disabling streaming.")
streamer = None
except Exception as e:
if sim_frame % 300 == 0:
logger.warning(f"[Stream] Frame error: {e}")
sim_frame += 1
if sim_frame % 30 == 0:
np.set_printoptions(precision=4, suppress=True, floatmode='fixed')
policy_obs = obs["policy"]
joint_pos = policy_obs["joint_pos"][0].cpu().numpy()
last_act = policy_obs["actions"][0].cpu().numpy()
if sim_frame == 30:
robot = env.unwrapped.scene["robot"]
jnames = robot.joint_names
print(f"\n{'='*70}")
print(f" ALL {len(jnames)} JOINT NAMES AND POSITIONS")
print(f"{'='*70}")
for i, name in enumerate(jnames):
print(f" [{i:2d}] {name:30s} = {joint_pos[i]:+.4f}")
arm_idx = [i for i, n in enumerate(jnames) if n.startswith("l_joint")]
print(f" Deduced left arm indices: {arm_idx}")
print(f"{'='*70}\n")
if not hasattr(env, '_arm_idx_cache'):
robot = env.unwrapped.scene["robot"]
jnames = robot.joint_names
env._arm_idx_cache = [i for i, n in enumerate(jnames) if n.startswith("l_joint")]
env._right_arm_idx_cache = [i for i, n in enumerate(jnames) if n.startswith("r_joint")]
arm_joints = joint_pos[env._arm_idx_cache]
right_arm_joints = joint_pos[env._right_arm_idx_cache]
try:
joy_dbg = teleop_interface.xr_client.get_joystick("left")
joy_str = f"[{joy_dbg[0]:+.2f}, {joy_dbg[1]:+.2f}]"
except Exception:
joy_str = "N/A"
if is_dual_arm:
eef_pos_left = policy_obs["eef_pos_left"][0].cpu().numpy()
eef_quat_left = policy_obs["eef_quat_left"][0].cpu().numpy()
eef_pos_right = policy_obs["eef_pos_right"][0].cpu().numpy()
print(f"\n---------------- [ROBOT STATE frame={sim_frame}] ----------------")
print(f"| Left Arm Joints (rad): {arm_joints}")
print(f"| Right Arm Joints (rad): {right_arm_joints}")
print(f"| Left EEF Pos (world, m): {eef_pos_left}")
print(f"| Right EEF Pos (world, m): {eef_pos_right}")
print(f"| Left Gripper Contact: {read_gripper_contact()}")
print(f"| Cmd left_arm(abs): {last_act[:7]}")
print(f"| Cmd wheel vel (rad/s): {last_act[7:11]}")
print(f"| Cmd left_gripper: {last_act[11:12]} (+1=open -1=close)")
print(f"| Cmd right_arm(abs): {last_act[12:19]}")
print(f"| Cmd right_gripper: {last_act[19:]} (+1=open -1=close)")
print(f"| Joystick (x=turn,y=fwd): {joy_str}")
print(f"----------------------------------------------------------------")
else:
eef_pos = policy_obs["eef_pos"][0].cpu().numpy()
eef_quat = policy_obs["eef_quat"][0].cpu().numpy()
if not hasattr(env, '_prev_eef_quat'):
env._prev_eef_quat = eef_quat.copy()
prev_q = env._prev_eef_quat
r_prev = R.from_quat([prev_q[1], prev_q[2], prev_q[3], prev_q[0]])
r_curr = R.from_quat([eef_quat[1], eef_quat[2], eef_quat[3], eef_quat[0]])
delta_rotvec = (r_curr * r_prev.inv()).as_rotvec()
delta_angle_deg = np.degrees(np.linalg.norm(delta_rotvec))
delta_axis = delta_rotvec / (np.linalg.norm(delta_rotvec) + 1e-9)
env._prev_eef_quat = eef_quat.copy()
approach_vec = r_curr.apply(np.array([1.0, 0.0, 0.0]))
print(f"\n---------------- [ROBOT STATE frame={sim_frame}] ----------------")
print(f"| Left Arm Joints (rad): {arm_joints}")
print(f"| EEF Pos (world, m): {eef_pos}")
print(f"| EEF approach vec (world): [{approach_vec[0]:+.3f}, {approach_vec[1]:+.3f}, {approach_vec[2]:+.3f}] (夹爪朝向)")
print(f"| EEF rot since last print: {delta_angle_deg:+.2f}° around [{delta_axis[0]:+.3f},{delta_axis[1]:+.3f},{delta_axis[2]:+.3f}] (world)")
print(f"| Cmd arm(abs pos+quat): {last_act[:7]}")
print(f"| Cmd wheel vel (rad/s): {last_act[7:11]}")
print(f"| Cmd gripper: {last_act[11:]} (+1=open -1=close)")
print(f"| Joystick (x=turn,y=fwd): {joy_str}")
print(f"----------------------------------------------------------------")
except Exception as e:
logger.error(f"Error during simulation step: {e}")
break
if streamer is not None:
streamer.close()
teleop_interface.close()
if is_dual_arm:
teleop_right_ref.close()
shared_client.close()
env.close()
if __name__ == "__main__":
main()
simulation_app.close()

View File

@@ -43,7 +43,7 @@ parser.add_argument("--drive_turn", type=float, default=1.5)
parser.add_argument("--stream-to", type=str, default=None, dest="stream_to")
parser.add_argument("--stream-port", type=int, default=12345, dest="stream_port")
parser.add_argument("--stream-bitrate", type=int, default=20_000_000, dest="stream_bitrate")
parser.add_argument("--debug-viewports", action="store_true", dest="debug_viewports", default=False)
parser.add_argument("--debug-viewports", action="store_true", dest="debug_viewports", default=True)
AppLauncher.add_app_launcher_args(parser)
args_cli = parser.parse_args()
@@ -95,7 +95,7 @@ def main() -> None:
drive_speed=args.drive_speed, drive_turn=args.drive_turn,
stream_to=args.stream_to, stream_port=args.stream_port,
stream_bitrate=args.stream_bitrate,
debug_viewports=args.debug_viewports,
debug_viewports=args.debug_viewports or bool(args.stream_to),
)
elif is_dual_arm:
print(f"[INFO] DualArmXrAgent (20D) | pos_sens={pos_sens} rot_sens={rot_sens}")

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Base teleoperation agent with main loop, env management, and reset handling."""

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Chassis (mobile base) control via XR joystick."""

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Periodic diagnostics reporter for teleoperation debugging."""

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Dual-arm XR teleoperation agent with chassis control."""
@@ -67,24 +67,48 @@ class DualArmXrAgent(BaseTeleopAgent):
policy_obs = obs["policy"]
robot = self.env.unwrapped.scene["robot"]
# Get robot root pose for world→body EEF conversion
root_pos_w = robot.data.root_pos_w[0].cpu().numpy()
root_quat_w = robot.data.root_quat_w[0].cpu().numpy()
# Read chassis
wheel_cmd, self._v_fwd, self._omega = self.chassis.get_commands(self.shared_client)
# Left arm
eef_pos_left = policy_obs["eef_pos_left"][0].cpu().numpy()
eef_quat_left = policy_obs["eef_quat_left"][0].cpu().numpy()
left_action = self.teleop_left.advance(current_eef_pos=eef_pos_left, current_eef_quat=eef_quat_left)
# Convert EEF observations from world frame to body frame
from .frame_utils import world_pose_to_root_frame
eef_pos_left_w = policy_obs["eef_pos_left"][0].cpu().numpy()
eef_quat_left_w = policy_obs["eef_quat_left"][0].cpu().numpy()
eef_pos_left_b, eef_quat_left_b = world_pose_to_root_frame(
eef_pos_left_w, eef_quat_left_w, root_pos_w, root_quat_w)
# Right arm
eef_pos_right = policy_obs["eef_pos_right"][0].cpu().numpy()
eef_quat_right = policy_obs["eef_quat_right"][0].cpu().numpy()
right_action = self.teleop_right.advance(current_eef_pos=eef_pos_right, current_eef_quat=eef_quat_right)
eef_pos_right_w = policy_obs["eef_pos_right"][0].cpu().numpy()
eef_quat_right_w = policy_obs["eef_quat_right"][0].cpu().numpy()
eef_pos_right_b, eef_quat_right_b = world_pose_to_root_frame(
eef_pos_right_w, eef_quat_right_w, root_pos_w, root_quat_w)
# Joint-locking: only convert when grip active
# XR controller works in body frame:
# - current_eef is body frame (converted above)
# - XR delta is treated as body-frame delta directly (no rotation)
# because in VR teleop the user's visual frame tracks the robot heading
# - output is body frame → send directly to IK
left_action = self.teleop_left.advance(current_eef_pos=eef_pos_left_b, current_eef_quat=eef_quat_left_b)
right_action = self.teleop_right.advance(current_eef_pos=eef_pos_right_b, current_eef_quat=eef_quat_right_b)
# Joint-locking: target is already in body frame, use directly
if self.teleop_left.grip_active or self._last_root_left is None:
self._last_root_left = convert_action_world_to_root(left_action, robot)[:7].clone()
self._last_root_left = left_action[:7].clone()
if self.teleop_right.grip_active or self._last_root_right is None:
self._last_root_right = convert_action_world_to_root(right_action, robot)[:7].clone()
self._last_root_right = right_action[:7].clone()
# Diagnostic: print once per second when left grip active
_cnt = getattr(self, '_diag_cnt', 0) + 1
self._diag_cnt = _cnt
if self.teleop_left.grip_active and _cnt % 30 == 0:
import math
print(f"[FRAME DIAG] root_quat_w={root_quat_w}")
print(f" eef_left_world=({eef_pos_left_w[0]:.3f},{eef_pos_left_w[1]:.3f},{eef_pos_left_w[2]:.3f})")
print(f" eef_left_body =({eef_pos_left_b[0]:.3f},{eef_pos_left_b[1]:.3f},{eef_pos_left_b[2]:.3f})")
print(f" ik_cmd_body =({self._last_root_left[0]:.3f},{self._last_root_left[1]:.3f},{self._last_root_left[2]:.3f})")
# left_arm(7) | wheel(4) | left_grip(1) | right_arm(7) | right_grip(1)
return torch.cat([

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Dual-arm + head tracking + trunk + stereo streaming XR agent."""

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Coordinate frame utilities for XR teleoperation."""

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Head tracking via XR headset (HMD) for robot head joint control."""

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""Sim-to-VR H264 stereo streaming manager."""
@@ -46,6 +46,8 @@ class StreamingManager:
def enabled(self) -> bool:
return self._enabled
_debug_saved = False
def send(self, scene, frame_count: int = 0) -> None:
"""Send one stereo frame from scene cameras. Auto-disables on failure."""
if not self._enabled:
@@ -53,6 +55,21 @@ class StreamingManager:
try:
left_rgb = scene["vr_left_eye"].data.output["rgb"][0].cpu().numpy()
right_rgb = scene["vr_right_eye"].data.output["rgb"][0].cpu().numpy()
# Save debug images once (frame 5 to ensure rendering is stable)
if frame_count == 150 and not self._debug_saved:
self._debug_saved = True
try:
from PIL import Image
import numpy as np
Image.fromarray(left_rgb[..., :3]).save("/tmp/vr_left_debug.png")
Image.fromarray(right_rgb[..., :3]).save("/tmp/vr_right_debug.png")
sbs = np.concatenate([left_rgb, right_rgb], axis=1)
Image.fromarray(sbs[..., :3]).save("/tmp/vr_sbs_debug.png")
print(f"[STREAM DIAG] Saved debug frames: left={left_rgb.shape}, right={right_rgb.shape}")
except Exception as e:
print(f"[STREAM DIAG] Save failed: {e}")
if not self._streamer.send_frame(left_rgb, right_rgb):
logger.warning("[Stream] Connection lost. Disabling streaming.")
self._enabled = False

View File

@@ -1,4 +1,4 @@
# SPDX-FileCopyrightText: Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# Copyright (c) 2025, Yutang Li, SIAT (yt.li2@siat.ac.cn)
# SPDX-License-Identifier: BSD-3-Clause
"""XR teleoperation controller with absolute IK target accumulation and grip clutch."""
@@ -50,6 +50,7 @@ class XrTeleopController:
self.grip_engage_threshold = 0.8
self.grip_release_threshold = 0.2
self.callbacks = {}
self._root_yaw_rad = 0.0 # current chassis heading, updated each frame
def add_callback(self, name: str, func: Callable):
self.callbacks[name] = func
@@ -63,6 +64,10 @@ class XrTeleopController:
self.target_eef_quat = None
self.require_grip_reengage = True
def set_root_yaw(self, yaw_rad: float):
"""Update the chassis heading so XR deltas are in body frame."""
self._root_yaw_rad = yaw_rad
def close(self):
if self._owns_client:
self.xr_client.close()
@@ -143,6 +148,16 @@ class XrTeleopController:
prev_xr_world_pos, prev_xr_world_quat_xyzw = transform_xr_pose(self.prev_xr_pos, self.prev_xr_quat)
world_delta_pos = (xr_world_pos - prev_xr_world_pos) * self.pos_sensitivity
# In VR teleop, user "inhabits" the robot: their physical forward (XR +X)
# maps directly to robot forward (body +X). No rotation needed — the VR view
# already rotates with the robot, so XR deltas naturally align with body frame.
# Diagnostic
if self.frame_count % 30 == 0 and self.target_eef_pos is not None:
print(f"[XR {self.arm}] delta=({world_delta_pos[0]:+.4f},{world_delta_pos[1]:+.4f},{world_delta_pos[2]:+.4f}) "
f"target=({self.target_eef_pos[0]:.3f},{self.target_eef_pos[1]:.3f},{self.target_eef_pos[2]:.3f})")
pos_norm = np.linalg.norm(world_delta_pos)
if pos_norm > self.max_pos_per_step:
world_delta_pos *= self.max_pos_per_step / pos_norm