Thanks for releasing RMBench — it's been great to work with.
I'm regenerating demonstrations locally with collect_data.py and hit a gap: two tasks that are present in the released dataset don't have environment definitions in the repo.
The discrepancy
- The HF dataset
TianxingChen/RMBench ships demo_clean demos for 12 tasks, including classify_blocks and storage_blocks (50 episodes each).
- But
envs/ in this repo contains only 10 task files — classify_blocks.py and storage_blocks.py are absent, and a repo-wide code search finds no reference to either. They're likewise missing from task_config/ and description/, and from the parent RoboTwin repo.
So these two appear to be data-only: the pre-generated episodes exist, but the code to (re)generate them isn't public.
Request
Would it be possible to release the environment files for classify_blocks and storage_blocks — the envs/*.py (scripted-expert + success check), plus any task-specific task_config / description entries?
I'd like to generate additional demonstrations with new seeds for these two tasks so they match the other ten; that isn't possible without the env code. Happy to contribute them back if that's easier.
Thanks!
Thanks for releasing RMBench — it's been great to work with.
I'm regenerating demonstrations locally with
collect_data.pyand hit a gap: two tasks that are present in the released dataset don't have environment definitions in the repo.The discrepancy
TianxingChen/RMBenchshipsdemo_cleandemos for 12 tasks, includingclassify_blocksandstorage_blocks(50 episodes each).envs/in this repo contains only 10 task files —classify_blocks.pyandstorage_blocks.pyare absent, and a repo-wide code search finds no reference to either. They're likewise missing fromtask_config/anddescription/, and from the parentRoboTwinrepo.So these two appear to be data-only: the pre-generated episodes exist, but the code to (re)generate them isn't public.
Request
Would it be possible to release the environment files for
classify_blocksandstorage_blocks— theenvs/*.py(scripted-expert + success check), plus any task-specifictask_config/descriptionentries?I'd like to generate additional demonstrations with new seeds for these two tasks so they match the other ten; that isn't possible without the env code. Happy to contribute them back if that's easier.
Thanks!