AnyCamVLA: Zero-Shot Camera Adaptation for Viewpoint Robust Vision-Language-Action Models
Paper • 2603.05868 • Published
AnyCamVLA — LVSM checkpoint (LIBERO-Plus custom)
Finetuned LVSM (scene decoder-only, 256px) checkpoint used in AnyCamVLA: Zero-Shot Camera Adaptation for Viewpoint Robust Vision-Language-Action Models (IROS 2026).
- Full finetune of the LVSM
scene_decoder_only_256base model on our custom LIBERO-Plus dataset (20k steps, 2 input views, no intrinsics augmentation). - Used for zero-shot camera adaptation: re-rendering perturbed camera observations back to the policy's training viewpoint.
- Optimizer state stripped; contains the
modelstate dict only (~1.3 GB).
Usage
Place the file in the AnyCamVLA repository:
hf download heo0224/AnyCamVLA-LVSM ckpt_0000000000020000.pt --local-dir LVSM/ckpt/LIBERO-Plus_custom/
Then follow the AnyCamVLA README to run LIBERO evaluation with --args.LVSM.
License
This checkpoint is a derivative of LVSM and is released under CC BY-NC-SA 4.0 (non-commercial), following the upstream LVSM license.
Links
- Code: https://github.com/heo0224/AnyCamVLA
- Paper: https://arxiv.org/abs/2603.05868
- Project page: https://heo0224.github.io/AnyCamVLA/
- Upstream LVSM: https://github.com/Haian-Jin/LVSM