It seems that the calculation of refline use the future information of ego traj. In this function “find_route”:

def find_route(traj, timestep, cur_pos, map_lanes, map_signals):
    lane_polylines = get_polylines(map_lanes)
    end_lane, end_point = find_map_waypoint(np.array((traj[-1].center_x, traj[-1].center_y, traj[-1].heading)), lane_polylines)
    cur_lane, _ = find_map_waypoint(cur_pos, lane_polylines)
    
    path_waypoints = []
    for t in range(0, len(traj), 10):
        lane, point = find_map_waypoint(np.array((traj[t].center_x, traj[t].center_y, traj[t].heading)), lane_polylines)
        path_waypoints.append(lane_polylines[lane][point])

    after_waypoints = []
    for waypoint in lane_polylines[end_lane][end_point:]:
        after_waypoints.append(waypoint)
    if len(after_waypoints) < 100:
        if map_lanes[end_lane].exit_lanes:
            lane = map_lanes[end_lane].exit_lanes[0]
            for waypoint in lane_polylines[lane]:
                after_waypoints.append(waypoint)

    waypoints = np.concatenate([path_waypoints, after_waypoints[::5]], axis=0)

    # generate smooth route
    tx, ty, tyaw, tc, _ = generate_target_course(waypoints[:, 0], waypoints[:, 1])
    ref_line = np.column_stack([tx, ty, tyaw, tc])
    
    # get reference path at current timestep
    current_location = np.argmin(np.linalg.norm(ref_line[:, :2] - cur_pos[np.newaxis, :2], axis=-1))
    start_index = np.max([current_location, 0])
    ref_line = ref_line[start_index:start_index+1000]

    # add speed limit and traffic signal info to ref route
    line_info = np.zeros(shape=(ref_line.shape[0], 1))
    speed_limit = map_lanes[cur_lane].speed_limit_mph / 2.237
    ref_line = np.concatenate([ref_line, line_info], axis=-1)
    signals = [Point([signal.stop_point.x, signal.stop_point.y]) for signal in map_signals[timestep].lane_states if signal.state in [1, 4, 7]]

    for i in range(ref_line.shape[0]):
        if any([Point(ref_line[i, :2]).distance(signal) < 0.2 for signal in signals]):
            ref_line[i, 4] = 0 # red light
        else:
            ref_line[i, 4] = speed_limit

    return ref_line
    

For the calculation of variable "path_waypoints", it seem that the future traj of ego is included.

There may be a problem with encoding the object type in the Class AgentEncoder.
You are using

type = self.type_emb(inputs[:, -1, 8].int())

And since object_type is one hot vector, I think the vector should be transformed into an integer first as:

object_type = (inputs[:, -1, 8:].argmax(dim=-1)+1)* inputs[:, -1, 8:].sum(dim=-1).int() type = self.type_emb(object_type)

Thank you for your amazing work about the interactive prediction and planning！

I have a detail I'd like to consult about joint prediction.
In joint prediction, is the probability of the final output scene-level, i.e. it has a dimension of 1 * num_mode?
Are joint predictions made in the same way as in DIPP?

I have this question because in The gameformer planner paper, you mentioned that "The model also provides probabilities associated with the different modes of each agent’s trajectory, resulting in tensors of shape N ×6. ", which confuses me a bit, seems to mean that each agent has a different probability output instead of dimension of 1×6.

Or could you please provide the code for the core model of Gameformer? My email is [email protected]. Thanks a lot!

I am very interested in your work and have similar thoughts. When running your code, in the data processing stage, there is a problem with training_20s, it cannot process the results.

for int_id in objects_of_interest:
interact_list.append(id_list[int_id])

An error will be reported in this command, and id_list is empty.

Hello MCZhi,
I appreciate your work and had similar ideas as yours, so I got very interesed in your work. I can't wait to see what you actually do in this code. I would be very appreciative if you could release the code soon.

Hello! I came across your paper and found it to be excellent – truly impressive and intriguing. If I can build on your work to achieve results, I will certainly cite your paper. However, I've encountered a problem at the moment: the installation of waymo-open-dataset-tf-2-11-0 is not successful. I'm getting the following error message:ERROR: Could not find a version that satisfies the requirement waymo-open-dataset-tf-2-11-0 (from versions: none)
ERROR: No matching distribution found for waymo-open-dataset-tf-2-11-0
I'm curious to know if this project is only compatible with Linux systems. I appreciate your response!Best regards.

how to use nuplan dataset for training and evaluation.

Hey!

Thanks for quite interesting work, I was going through the project website and its quite appealing. I am not sure if you already have preprint available, but currently the arxiv link is point to old paper.

Hope you have a nice week!

What are the hyperparameters used for the best model on the Waymo Interaction Prediction challenge?
The one reported to have a minADE of 0.9161.

There are some discrepancies between the default code values and the ones reported in the paper appendix.
For example, do you consider 20 or 32 neighbors in the preprocessing code for the best results?

Also if you can share the model weights that would be very useful. Thanks :)