Hi, thank you for your great work and for releasing the codebase. While carefully reading Section 3.3 and comparing it with the implementation, I found several points that I could not clearly verify in the current code. I may be misunderstanding some parts, so I would appreciate any clarification.
1. Three separate self-attention mechanisms for individual tasks
In the paper, the Collaborative Interaction Module is described as incorporating three separate self-attention mechanisms, each dedicated to an individual task. From my understanding, these correspond to agent queries (det), map queries, and planning queries. However, in the released configuration the SeparateAttention module is defined with separate_list = [["det"], ["map"]], which seems to create separate self-attention only for det and map queries. I could not find a separate self-attention branch dedicated to planning queries. Instead, planning queries appear to participate through the inter_graph_model. Could you clarify whether the planning query is intentionally excluded from the per-task separate self-attention, or if there is another configuration where planning queries also have their own separate self-attention?
2. Distance-based attention
The paper describes a geometric attention mechanism that incorporates distance-based interactions between map-agent, map-map, and agent-map anchors. While reviewing the code, I noticed that the distance attention mechanism appears to be implemented (for example in get_distance_attn_mask and distance_tau), but in the released configuration it seems to be disabled by default through with_distance_attn_mask=False. Because of this, it appears that the distance-based attention might not be active in the default setup. Could you clarify whether with_distance_attn_mask=True was used for the experiments reported in the paper?
3. No direct information exchange between map and agent (det) queries
The paper mentions interactions such as map-agent, map-map, and agent-map within the Collaborative Interaction Module. However, while reviewing the attention modules in the configuration, it seems that det and map queries are processed independently in SeparateAttention, and planning queries attend to det and map in InteractiveAttention. In the current configuration, I could not find a module where map queries attend to agent queries, agent queries attend to map queries, or map queries attend to other map queries within the collaborative interaction stage. I may be misunderstanding the intended interaction structure, so I would appreciate clarification on whether these interactions are expected to be enabled through a different configuration.
Thank you again for making the code available. Any clarification would be very helpful for understanding the implementation.
Hi, thank you for your great work and for releasing the codebase. While carefully reading Section 3.3 and comparing it with the implementation, I found several points that I could not clearly verify in the current code. I may be misunderstanding some parts, so I would appreciate any clarification.
1. Three separate self-attention mechanisms for individual tasks
In the paper, the Collaborative Interaction Module is described as incorporating three separate self-attention mechanisms, each dedicated to an individual task. From my understanding, these correspond to agent queries (det), map queries, and planning queries. However, in the released configuration the SeparateAttention module is defined with separate_list = [["det"], ["map"]], which seems to create separate self-attention only for det and map queries. I could not find a separate self-attention branch dedicated to planning queries. Instead, planning queries appear to participate through the inter_graph_model. Could you clarify whether the planning query is intentionally excluded from the per-task separate self-attention, or if there is another configuration where planning queries also have their own separate self-attention?
2. Distance-based attention
The paper describes a geometric attention mechanism that incorporates distance-based interactions between map-agent, map-map, and agent-map anchors. While reviewing the code, I noticed that the distance attention mechanism appears to be implemented (for example in get_distance_attn_mask and distance_tau), but in the released configuration it seems to be disabled by default through with_distance_attn_mask=False. Because of this, it appears that the distance-based attention might not be active in the default setup. Could you clarify whether with_distance_attn_mask=True was used for the experiments reported in the paper?
3. No direct information exchange between map and agent (det) queries
The paper mentions interactions such as map-agent, map-map, and agent-map within the Collaborative Interaction Module. However, while reviewing the attention modules in the configuration, it seems that det and map queries are processed independently in SeparateAttention, and planning queries attend to det and map in InteractiveAttention. In the current configuration, I could not find a module where map queries attend to agent queries, agent queries attend to map queries, or map queries attend to other map queries within the collaborative interaction stage. I may be misunderstanding the intended interaction structure, so I would appreciate clarification on whether these interactions are expected to be enabled through a different configuration.
Thank you again for making the code available. Any clarification would be very helpful for understanding the implementation.