Failed To Train the AnymalTerrain Task
AlanSunHR opened this issue · 0 comments
AlanSunHR commented
Hi, recentlty I tried to train the AnymalTerrain task using the latest codes in this repo and the lastest version of Isaac Sim.
Without changing any parameters, I tried 3 seeds, but the Anymal robots just stand still all the time in these 3 seeds. I tried to play with the demo and load the checkpoint from the NVidia server, it worked perfectly well.
Another problem is that the GPU util is pretty low (around 20% on a 4080 GPU), I also trained the same task using the Isaac Sim version of 2022.2.1, the GPU util is about 60% and the overall training speed is 5x faster.
So my question is:
- Are the current parameters for AnymalTerrain task robust enough for different random seeds?
- Do I have to design a curriculum to successfully train AnymalTerrain (lower penalty/noise at first and then increase the penalty)?
- Why are the performance of different versions of Isaac Sim so different?
Thanks!