jhlee0619's picture
Fine-tuned world models
66a1264 verified
Raw History Blame Contribute Delete
1.23 kB
Fine-tuned world model released with Unstick (https://github.com/jhlee0619/unstick)
by Junhyeok Lee, Dongheon Lee, Kyu Sung Choi, under the MIT License of that repository.
This model is the on-policy fine-tuning (FT, scripts/finetune.py) of a released checkpoint:
pushobj_shape_shift, pusht_visual_shift, mediummaze_dynamics_shift, diversemaze_layout_shift: AdaJEPA
(https://github.com/agentic-learning-ai-lab/adajepa, MIT License, Copyright (c) 2025 gaoyuezhou)
dinowm_pusht: DINO-WM (https://github.com/gaoyuezhou/dino_wm, MIT License, Copyright (c) 2025 gaoyuezhou)
tsdino_pusht: Temporal Straightening (https://github.com/agentic-learning-ai-lab/temporal-straightening,
MIT License, Copyright (c) 2025 gaoyuezhou)
The DINOv2-based models (dinowm_pusht, tsdino_pusht) contain DINOv2 ViT-S/14 weights
(https://github.com/facebookresearch/dinov2, Apache License 2.0).
The full license texts are in LICENSES/ of the repository.
Fine-tuning changed only the tensors that scripts/finetune.py trains (those that AdaJEPA adapts; predictor only for
tsdino_pusht). hydra.yaml is the released model's configuration, copied unchanged.
diversemaze_layout_shift is fine-tuned with --norm-stats val, the default of finetune.py.