DP3 baseline not generalizing.

[dennisushi]

Hi, I am training the model in the simplest possible real task: reaching an object. This is done with 7DoF Panda controlled via end effector pose (world-frame XYZ-RPY 6 actions), training SimpleDP3 on 40 demonstrations (with default 0.02 validation ratio as in your configs) until the test score converges near 0 (about 2k epochs).

As you can see in the plot, one of the trajectories (which I believe is assigned as a validation trajectory) is incorrect. My main suspicion is that my data is non-Markovian as I do it continuously with real-time teleoperation and that the model cannot handle that (e.g. short pauses in the movement). If my suspicion is correct, I suspect a larger time horizon may fix it partially.

Is the test score based on the validation trajectories as it doesn't match what I observe? Is this overfitting to demonstrations normal for the model?

[YanjieZe]

Hi, thank you for your interest!

Your experiments are actually interesting. As I observe in the figure, the line that you mention as not right, seems to be correct in its earlier part, but goes into messy in the end. Right? I am not pretty sure whether it is showing "not generalizing". Instead, I think it is more like that that trajectory is hard to predict (a sudden twist).

I think one property of diffusion polices (DP, DP3, and other variants) is to make the trajectory prediction smooth. The phenomenon you see might be caused from this property.

[dennisushi]

The bad predictions start much before the twist (see rightmost part of rightmost plot), and that twist isn't sudden in time, just in space. It seems to be more of a generalization issue, as I do not observe task-oriented behaviour when new data is presented - which means either the data is insufficient or in some incorrect format.

[kingchou007]

Hi @dennisushi

Did you evaluate dp3 in real world? I've tried implementing dp3 in real-world task, but I'm struggling to achieve any meaningful results.

I don't know what am I missing or doing wrong that is leading to this extremely low success rate?

[dennisushi]

@kingchou007 Hi, I tried to evaluate in real world. The behaviour I am getting is consistently bad. If you want we can chat and figure out where we are failing together.

[kingchou007]

@kingchou007 Hi, I tried to evaluate in real world. The behaviour I am getting is consistently bad. If you want we can chat and figure out where we are failing together.

Hi @dennisushi,

Sure, I've sent an e-mail to your Gmail.