Play2Perfect checkpoints

State-based policy checkpoints for Play2Perfect (arXiv: 2606.26428).

  • play/model.pth β€” Stage-1 "play" pretrained policy.
  • tight_insertion/model.pth β€” Stage-2: tight insertion (L-peg, 0.5 mm).
  • beam_assembly_step1/model.pth β€” Stage-2: beam assembly step 1.
  • beam_assembly_step2/model.pth β€” Stage-2: beam assembly step 2.
  • screwing/model.pth β€” Stage-2: screwing (furniture leg).
  • fork_insertion/model.pth β€” Stage-2: YCB fork into a flatware holder (handle first, 1 mm clearance).
  • plug_insertion/model.pth β€” Stage-2: USB-C wall plug into a power-strip socket (0.5 mm clearance).

The checkpoints above are for the KUKA iiwa 14 + Sharpa hand (task Isaacsimenvs-PreciseAssembly-Direct-v0).

Flexiv Rizon 4s + right Wuji Hand 2

Under flexiv_wuji_right_human_like/ (task Isaacsimenvs-PreciseAssemblyFlexivWujiRight-Direct-v0):

  • play/model.pth β€” Stage-1 "play" pretrained policy.
  • tight_insertion/model.pth, beam_assembly_step1/model.pth, beam_assembly_step2/model.pth, fork_insertion/model.pth, plug_insertion/model.pth β€” Stage-2 policies finetuned from it (domain randomization off).

There is no Flexiv screwing policy yet. These are snapshots of runs that are still training and may be updated.

Download

From the Play2Perfect repo:

python download_checkpoints.py                                       # KUKA + Sharpa
python download_checkpoints.py --robot flexiv_wuji_right_human_like  # Flexiv + Wuji
Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading