Questions about evaluation setting and delay time behavior

[junsoo-lim-kaist]

Hello thank you very much for sharing this impressive work!
I was especially surprised to see that the demo extends all the way to fast reaction tasks such as playing ping pong which was very impressive.

I have a few questions regarding the evaluation in Figure 6 and Table 1.

Execution horizon setting in Figure 6
In the left plot of Figure 6, it seems that for the Kinetix tasks faster reaction generally leads to a higher solve rate. In the right plot ,since the execution horizon is defined as s = max(d, 1), it looks like the decrease in success rate may be influenced by both s and d increasing together. If the goal is to demonstrate that the proposed method resolves misalignment caused by delay, I was wondering why s was not fixed to a constant value.
Is there a specific reason for choosing s = max(d, 1)? Was this choice mainly made to follow the experimental setup of prior work or does this setting have an additional motivation?

Action quantization in Figure 6
Are the results in Figure 6 obtained with action quantization enabled or are these results measured without applying action quantization?

Completion time trend in Table 1
In Table 1 the completion time decreases as the inference delay increases.
I was wondering if you could briefly explain the reason behind this behavior?

Thank you very much for your time and for making the code and paper available.

[Sakits]

Hi @junsoo-lim-kaist ,

Thank you so much for your interest in our work, and apologies for the late reply!

1. Regarding the choice of setting s = \max(d, 1): yes, this was done to follow the setting used in RTC. It corresponds to a scenario where the execution horizon fully covers the inference delay.
2. No action quantization was applied for the results shown in Figure 6.
3. This is because the 4090 GPU has an inference latency of ~3 steps (as mentioned in the figure caption). When the delay is too small, the execution window during inference cannot cover the full inference delay, resulting in stall. Even though a smaller delay yields more precise actions and fewer steps, it ends up taking longer due to waiting on inference.

Hope these address your questions! :)

Best,
Jiaming

[junsoo-lim-kaist]

Hi Jiaming,

Thank you for the detailed and clear explanation. That fully addresses my questions.
I really appreciate you taking the time to clarify this, and thanks again for sharing such insightful work.

Best regards,
Junsoo