Clarification: does the LogoPlanner startgoal Isaac simulation evaluation use simulator GT pose for MPC tracking?

[mhqin1018]

Hi, thanks for releasing LoGoPlanner and the NavDP benchmark.

I want to clarify the simulation evaluation protocol. The paper/project page describes LoGoPlanner as using implicit localization without an
external localization/odometry module. From the released code, I agree that the LoGoPlanner policy/server itself does not seem to take GT pose
as input.

Policy/server side

logoplanner_server.py only receives goal/RGB/depth and returns trajectories:

goal_data = json.loads(request.form.get('goal_data'))
image_file = request.files['image']
depth_file = request.files['depth']

execute_trajectory, all_trajectory, all_values, trajectory_mask, sub_pointgoal_pd = \
    logoplanner_navigator.step_pointgoal(goal, image, depth)

return jsonify({
    'trajectory': execute_trajectory.tolist(),
    'all_trajectory': all_trajectory.tolist(),
    'all_values': all_values.tolist(),
    'sub_pointgoal_pd': sub_pointgoal_pd.tolist()
})

Isaac simulation evaluator side

In eval_startgoal_wheeled.py, the low-level MPC tracker appears to use simulator GT pose as feedback. x0 is built from Isaac camera/root state:

camera_pos = env.unwrapped.scene.sensors['camera_sensor'].data.pos_w.cpu().numpy()
camera_rot_quat = env.unwrapped.scene.sensors['camera_sensor'].data.quat_w_world.cpu().numpy()
camera_rot = R.from_quat(camera_rot_quat[:, [1, 2, 3, 0]]).as_matrix()

robot_vel = env.unwrapped.scene.articulations['robot'].data.root_lin_vel_w[0, :2].norm().cpu().numpy()
robot_ang_vel = env.unwrapped.scene.articulations['robot'].data.root_ang_vel_w[0, 2].cpu().numpy()

x0 = np.stack([
    camera_pos[:, 0],
    camera_pos[:, 1],
    np.arctan2(camera_rot[:, 1, 0], camera_rot[:, 0, 0]),
    [robot_vel],
    [robot_ang_vel],
], axis=-1)

Then this state is passed to MPC:

opt_u_controls, opt_x_states = mpc.solve(x0[i, :3])

And tracking_utils.py sets it as the MPC initial state:

self.opti.set_value(self.opt_x0, x00)

My understanding

1. LoGoPlanner policy/server does not use GT pose as input.
2. The released Isaac simulation evaluator does use simulator GT pose for low-level MPC tracking.
3. This GT feedback may correct tracking/odometry errors during simulation, even though it is not part of the policy input.

Could you clarify whether the reported simulation results use this GT-pose MPC feedback？
Thanks.