Views
No views yet
1from posevla.modeling_posevla import PoseVLAPolicy, PoseVLAConfig, bin_tokenizer
2from utils.mapping_token import decode_text_to_scene_with_tokenizer
3
4# Load model
5policy = PoseVLAPolicy.from_pretrained("hetolin/PoseVLA-stage1", local_files_only=False, config=posevla_config)
6policy = policy.eval().to(torch.bfloat16).cuda()
7
8# Inference
9output_res = policy.forward_evaluate_ntp(batch)
10pred_text = output_res["pred"][0]
11pred_res = decode_text_to_scene_with_tokenizer(pred_text, bin_tokenizer)infer_grounding3d.py for complete real-world RGB-D inference pipeline.| Hyperparameter | Value |
|---|---|
| Base model | PaliGemma-3B-pt-224 |
| Action Expert | π0 (Flow Matching, from scratch) |
| Image resolution | 224 × 224 |
| Optimizer | AdamW |
| Learning rate | 5e-5 |
| Weight decay | 1e-10 |
| Precision | bf16 |
| GPUs | 16 × H20 |
| Batch size | 7 per GPU |
| Training steps | 100K |
1@article{lin2026posevla,
2 title={PoseVLA: Universal Pose Pretraining for Generalizable Vision-Language-Action Policies},
3 author={Lin, Haitao and Yu, Hanyang and Huang, Jingshun and Zhang, He and Ling, Yonggen and Tan, Ping and Xue, Xiangyang and Fu, Yanwei},
4 journal={arXiv preprint arXiv:2602.19710},
5 year={2026}
6}