Foot Position Prior : Sequence or final position

[Batou1406]

Foot Position Prior : Sequence or final position

The foot position prior is where the foot should land after the swing trajectory. It is outputed by the neural network.

Now it is outputed as a sequence of shape (batch_size, num_legs, 3, time_horizon). This means only one of the positions in the sequence is actually used. It is the position when the foot enters in contact with the ground (ie. the contact in the contact sequence c change from 0 to 1).

Question : Should we keep it as a sequence or change it to final position only ?

• Pros A sequence allow the controller to define different behaviour of the leg depending on the time horizon and the contact sequence.
• cons It greatly increase the learning space and thus makes the training more difficult. Moreover RL controller is already informed of the contact sequence (it has leg frequency and duty cycle) and thus could infer it. Which mitigates the pros...

[Batou1406]

reduce p from a sequence to two or three touch down position. Only the first one would be used as long as there are no sampling controller in the loop. Then, they would be used for the futur prediction.

[Batou1406]

Changed from a time sequence to a number of prediction : shape(batch_size, num_legs, 3, time_horizon) to shape(batch_size, num_legs, 3, number_predict_step)

this should greatly reduce the learning space, while giving several different touch down position to the sampling controller if needed.

Closed for now