Hi FlashSAC team, thanks for the great work! :)
I'm trying to integrate FlashSAC into mjlab for a humanoid velocity-tracking task. Reward terms are ported from the PPO recipe, and I've already rebalanced several weights following the guidance in Table 14 of the paper.
Across multiple runs, the policy still converges to a degenerate behavior: the robot stands almost still and only tilts its torso slightly toward the velocity command, instead of walking.
A few questions:
- Is this a known failure mode? Any specific reward terms that typically need the most attention beyond what's listed in the paper?
- Is there a recommended reward template for mjlab specifically ?
Happy to share my reward config and training curves. Thanks!
Hi FlashSAC team, thanks for the great work! :)
I'm trying to integrate FlashSAC into mjlab for a humanoid velocity-tracking task. Reward terms are ported from the PPO recipe, and I've already rebalanced several weights following the guidance in Table 14 of the paper.
Across multiple runs, the policy still converges to a degenerate behavior: the robot stands almost still and only tilts its torso slightly toward the velocity command, instead of walking.
A few questions:
Happy to share my reward config and training curves. Thanks!