Skip to content

Conversation

@linden713
Copy link

Fixes an unreasonable terrain-difficulty downgrade behavior in the terrain curriculum. The community has reported similar issues where the curriculum (which measures distance from the env origin) can fail in some cases. Ref: IsaacLab discussion.

What I changed

  • Add params to control upgrade and downgrade
  • Prevent downgrade when the episode end with timed out.
  • Consider only the last command window when estimating expected distance. Episode length is 20s and commands are resampled every 10s. This prevents agents that are slow early but fast later from being unfairly downgraded.

Why

Even for flat terrains, the level never rises above 1.5, which is unreasonable and prevents normal progression.

This image shows terrain levels on flat terrain — it should reach 6, but the original function is capped at 1.5. Staying at 1.5 on flat terrain is unreasonable; reaching 6 is normal and expected.
81ba1ffdb3b7c972693355e8750687fb

This image shows results from my go2 experiments on mixed terrains (stairs, slopes, etc.) using fixed terrain level function. The curriculum reaches 6, which matches expected behavior.
6d3952e4-97d9-47c3-88ec-f8a7c8135afe

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant