lang_tokens: (1, 77, 4096) T5-XXL 语言嵌入 (示例使用随机嵌入演示;真实部署可加载 google/t5-v1_1-xxl 编码)img_tokens: (1, 4374, 1152) SigLIP 视觉 token (2帧历史 × 3相机 × 729 patches, 来自 google/siglip-so400m-patch14-384)state_tokens: (1, 1, 128) 统一动作空间 proprio (含双臂14维+mask)ctrl_freqs: [25] Hzaction_chunk (1, 64, 128) 统一空间连续动作,经 state_adaptor 反归一化可得 14 维关节动作torch.npu.is_available()=True, device_count=2, device_name=Ascend910_9362torch.npu.synchronize()scripts/validate_policy_outputs.py --atol 1e-4 --rtol 1e-3scripts/benchmark_policy.py --latencies-ms latencies.json --items-per-call 64 --unit actionsCANN 8.5.1, torch 2.9.0+cpu, torch_npu 2.9.0, timm, diffusersSiglipVisionTower 于 NPUpip install -r requirements.txt
python inference.py

