Default Collision Detection for Manipulation Example Inaccurate

[An-Axolotl]

Hello,

I recently ran your allegro_reorient example. If you look closely at the resulting policy the ball is bounced between fingers without accurate collision rendering between the ball and the fingers. This can be seen more clearly by just slightly changing the reward function.

Altered Reward Function:
I altered the reward function to reward multi-finger contact and a small penalty for the ball bouncing on top of the fingers. The result is a policy which relies on the ball interfering and resting where the hinge of the allegro finger is. You can see how the collision detection seems to really fail here as it rests here for a bit with no rebound.

RewardSearch_HeightPenalty+MultiFingerContactRewardAllegro-2025-06-14_01.42.43.mp4

The allegro model provided seems to be have altered by the authors to exclude collisions of fingers and palm from the original. Adding these back in causes memory issues.

Please let me know how this can be fixed. The default example + default reward also have this collision issue but on a less noticeable scale.