DRAM Overflow Occurs After Increasing Training Steps in RSL-RL

[H-Hisamichi]

Hello everyone,

I am using a Hexapod robot to perform reinforcement learning on complex tasks.
After finishing the design of the command and reward functions, I increased the max_iterations value to allow the policy to fully converge. However, at step 4010, a DRAM overflow occurred.

Are there any recommended ways to reduce DRAM usage during RL training?

Thank you!

Isaac Sim: 4.5.0
Isaac Lab: 2.0.2
DRAM: DDR4 64GB

[HaFred]

Any flushing you did in between to mitigate?

[H-Hisamichi]

Hello @HaFred,

No, I haven't done any DRAM flushing.
All memory-related settings are left as default.

[H-Hisamichi]

Hello everyone,

I decided to upgrade the memory.
I simplified the mesh as much as possible, but the problem wasn't resolved.