Dynamic in task Dump bin bigbin

[RuiWang724]

Hi,

Thanks for your excellent work on DOMINO!

We ran evaluation tasks on the DOMINO benchmark and found that in the dump_bin_bigbin task, in /envs/dump_bin_bigbin.py, line 97 sets c.set_kinematic(True), which causes the balls to move together with the trash bin. However, their kinematic state is not released at the end. As a result, when the robot arm tilts the small bin to pour the balls into the large bin, the balls do not actually fall out.

In our own experiments, the success rate of dump_bin_bigbin was consistently 0. We also noticed that in the paper “Towards Generalizable Robotic Manipulation in Dynamic Environments”, the reported success rate for dump_bin_bigbin is also 0 across all dynamic experiments.

We tried adding a fix that releases the balls at the end of the workflow. The code is available here: dump_bin_bigbin_fix. Would it be acceptable to modify /DOMINO/envs/dump_bin_bigbin.py as in the linked code?

Thank you very much for your help. We look forward to your reply!

[FangHeng]

Thanks @RuiWang724 for the clear report and for sharing the reference implementation.

I confirmed that this is a real issue in the dynamic policy-eval path: the garbage spheres are initialized as kinematic so they follow the moving small bin, but stop_dynamic_object_motion() only stopped the small bin and did not release the spheres back to physics. The expert play_once() path already released them after the dynamic workflow, so the missing piece was the eval contact-stop path.

I applied a minimal fix in dump_bin_bigbin: added an idempotent release_garbage_physics(), call it after stop_dynamic_object_motion() succeeds, and reuse the same helper in the expert dynamic path. check_success() remains unchanged.

Thanks again for catching this and proposing a clean fix.

[RuiWang724]

Hi @FangHeng,

Thanks for the quick fix and for updating the code!

I used the latest code, and the dump_bin_bigbin task can now be completed successfully.

Best regards