closed_loop.py: should saved frame metadata use planning-time current_info or post-step info

[DanqiZhao21]

Question about frame serialization semantics in closed_loop.py

I am trying to understand the intended meaning of each saved frame in data.pkl.

In the original implementation of closed_loop.py, the code does:

acc, steer_rate = traj2control(plan_traj, info)
obs, reward, terminated, truncated, info = env.step(action)

imu_plan_traj = plan_traj[:, [1, 0]]
imu_plan_traj[:, 1] *= -1
global_traj = traj_transform_to_global(imu_plan_traj, info['ego_box'])

save_data['frames'].append({
    'time_stamp': info['timestamp'],
    'ego_box': info['ego_box'],
    'obj_boxes': info['obj_boxes'],
    'planned_traj': {
        'traj': global_traj,
        'timestep': 0.5
    },
    'collision': info['collision'],
    'rc': info['rc']
})

My understanding

This means the saved frame uses post-step info for everything:

• time_stamp
• ego_box
• obj_boxes
• planned_traj
• collision
• rc

But the plan_traj itself was generated from the observation/info before env.step().

So semantically, this looks like:

1. generate plan_t from obs_t
2. convert plan_t to control
3. execute one step to get info_{t+dt}
4. save the frame using info_{t+dt} and also transform the original plan using info['ego_box']

What confuses me

This seems to make the saved trajectory no longer correspond exactly to the state from which it was generated.

In other words, the recorded frame is not a pure planning-time snapshot, but rather something like:

• plan generated at time t
• saved together with state/outcome at time t+dt

I want to confirm whether this is intentional.

Specific questions

1. Is this the intended semantics of a frame in data.pkl?

   • a post-step frame that stores the previous plan aligned to the new ego state?

2. Was it intentional to transform plan_traj using info['ego_box'] after env.step(), instead of the ego pose at planning time?

3. Is collision / rc supposed to be interpreted as:

   • labels for the saved frame after execution, or
   • labels for the state when the plan was generated?

4. If this is intentional, what is the reasoning behind this choice?

   • for example, is the goal to evaluate the plan after one control interval has been consumed?

Why I am asking

I noticed that using post-step info for the saved frame can change evaluation behavior noticeably, and I want to make sure I follow the original intended protocol rather than reinterpreting it incorrectly.

if plan_traj is not None:
    current_info = info  # info before env.step()

    acc, steer_rate = traj2control(plan_traj, info, wheelbase=wheelbase)

    action = {'acc': acc, 'steer_rate': steer_rate}
    obs, reward, terminated, truncated, info = env.step(action)  # info after env.step()

imu_plan_traj = plan_traj[:, [1, 0]]
imu_plan_traj[:, 1] *= -1

global_traj = traj_transform_to_global(imu_plan_traj, current_info['ego_box'])

save_data['frames'].append({
    'time_stamp': current_info['timestamp'],
    'ego_box': current_info['ego_box'],
    'obj_boxes': current_info['obj_boxes'],
    'planned_traj': {
        'traj': global_traj,
        'timestep': 0.5
    },
    'collision': info['collision'],
    'rc': info['rc']
})
’’’
if this change is right?
Thanks.