code_and_weights · 已核验官方资源
审核码：verified_official; 官方关系：apache_code_and_libero_checkpoint_with_independent_version_sensitive_reproduction; 最近核验：2026-07-29代码：Python package with π0/π0.5 training, inference, LoRA, benchmark and LeRobot adapters
权重：official 4B π0.5 LIBERO async5 checkpoint exists; SmolVLA checkpoint is not released
数据：uses external LeRobot/LIBERO datasets; Kinetix release remains incomplete
许可证：Apache-2.0_code_checkpoint_not_disclosed；Apache-2.0 covers repository source; the Hugging Face checkpoint has no model card or artifact-specific license metadata.
依赖：Python 3.10、PyTorch、torchvision、torchcodec、LeRobot datasets/models、CUDA Graphs for low-latency path
硬件/传感器：RTX 4090/5090 or H100 measured、Galaxea R1 Lite、LeRobot SO-101
实机证据：Three manipulation tasks with 20 rollouts per method; ping-pong 11/20 and whack-a-mole five rounds.
独立复现：Issue #22 independently reproduced LIBERO after pinning MuJoCo 3.3.7 and NumPy 1.24.4; high-delay result changed from 74.8% to 86.2%.
限制/反面证据：Strong MuJoCo and NumPy version sensitivity；Paper LIBERO numbers use one seed；Some checkpoints and evaluation data missing；HF checkpoint license missing
当前建议：实时VLA主流技术2.0·社区已复现；Start from the reproduced MuJoCo/NumPy pins and benchmark latency plus closed-loop accuracy together.
重点标签：real-time-vla、code:verified、weights:partial、community:positive-mixed
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4
