What does SAIL refer to?

[chenzhutian]

Thanks for your wonderful repo!
May I ask what does SAIL policy refer to?
Based on the paper I only find SARL but not SAIL. Is SAIL an upgrade version of SARL?
Thanks!

[YuejiangLIU]

Thanks for your interest in our work.

The SAIL refers to socially attentive imitation learning, a model used in our Social-NCE paper.

The key differences between SAIL and SARL proposed in our CrowdNav paper are the following,

• SAIL is a policy net that directly outputs an action, whereas SARL is a value net with a discrete action space.
• In addition to the action, SAIL also outputs the encoded feature vector, which is robustified by our proposed social contrastive loss during training.
• SAIL contains fewer parameters than the SARL. We empirically find that removing some dense MLPs makes the SAIL easier and faster to train, without degrading the performance.

[chenzhutian]

Awesome! Thanks for your explanation and will look into the details.
I am currently working on a project based on your works. It is not for robots but for user interfaces, though.
I am also wondering have you tried the GA3C method mentioned in [1]?

Thanks!

[1] Michael Everett‡ , Yu Fan Chen† , and Jonathan P. How‡. Motion Planning Among Dynamic, Decision-Making Agents with Deep Reinforcement Learning.

[chenzhutian]

Another thing is that I would like to remove the imitation in SAIL and only use it to control an agent (eg, following a target while avoiding collision), is this possible?

Thanks again~

[chenzhutian]

One more question about SAIL 😃 -
it seems the state is also different from SARL? The robot state is (px, py, vx, vy, gx, gy) and the human state is (px, py, vx, vy). Are these in the robot center coordination? I can find the code (eg, the rotate function used in SARL) to transfer the coordination.

Best

[YuejiangLIU]

Awesome! Thanks for your explanation and will look into the details. I am currently working on a project based on your works. It is not for robots but for user interfaces, though. I am also wondering have you tried the GA3C method mentioned in [1]?

Thanks!

[1] Michael Everett‡ , Yu Fan Chen† , and Jonathan P. How‡. Motion Planning Among Dynamic, Decision-Making Agents with Deep Reinforcement Learning.

As far as I remember, the GA3C was actually one of the baselines in our CrowdNav paper, where we named it as LSTM-RL in order to distinguish its core LSTM module from the attention mechanism used in ours

[YuejiangLIU]

Another thing is that I would like to remove the imitation in SAIL and only use it to control an agent (eg, following a target while avoiding collision), is this possible?

Thanks again~

It depends on your use case (e.g., environment).

If your environment is very similar to ours, I believe, you can try and use an already trained SAIL. Else if there is a clear domain shift, I'm afraid the model has to be adapted accordingly.

[YuejiangLIU]

One more question about SAIL 😃 - it seems the state is also different from SARL? The robot state is (px, py, vx, vy, gx, gy) and the human state is (px, py, vx, vy). Are these in the robot center coordination? I can find the code (eg, the rotate function used in SARL) to transfer the coordination.

Best

If I remember correctly, they are global coordinates, but I'm not 100% sure. I cannot recall some details, unfortunately. Maybe it's better to verify that by running the code till that line.

[chenzhutian]

Great, I will run the code and prob into the details. Thanks!

[YuejiangLIU]

I'm closing the issue for now. Please feel free to open a new one, if you encounter any other questions later on.