RARA – pre-trained Diffusion Policy (RoboTwin 2.0)
pretrain_ws2k.pt: Diffusion Policy (ResNet-18 head + wrist camera encoders, UNet denoiser, official RoboTwin
hyper-parameters) trained for 60 epochs (batch 256) on 2000 RoboTwin 2.0 demonstrations of five tasks
(ShuaKang/my_roboTwin2.0_training, dp_zarr_ws/).
It is the starting point for FT, Co-FT and RARA in github.com/Shua-Kang/RARA.
Format: EMA weights + training config (load_policy in rara/workspace.py).
bash scripts/download.sh # in the RARA repository; places this file in data/weights/