凤
weixin_44093752/MidoriChou_ai_final_12_act_2_Ascend_NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

MidoriChou/ai_final_12_act_2 Ascend NPU adaptation

This delivery keeps the official LeRobot ACT checkpoint and runs the official ACT network on Ascend910 npu:0. The model takes two RGB camera views with shape 3x480x640 and a 9-dimensional robot state, then returns an action chunk with shape 100x8.

Official source

  • Hugging Face model: https://huggingface.co/MidoriChou/ai_final_12_act_2
  • Revision: 92654eb0e3f9ee4329d54a599165aeb12512e92f
  • Official runtime: https://github.com/huggingface/lerobot/tree/v0.3.3
  • Checkpoint: model.safetensors, 206,720,224 bytes
  • SHA-256: 9da5b06966343936951d917de0225b07db1ca40eba100f841918d4ef389a223d

NPU smoke result

The checkpoint loaded strictly with 234 state keys and zero missing or unexpected keys. The model has 51,602,312 parameters. On a real Ascend910 npu:0, the delivery script completed the official input contract in 0.16056 seconds. The output was float32, shape [1, 100, 8], and all values were finite. This is an NPU smoke gate only; it does not claim full robotics-task accuracy or a CPU/NPU precision pass.

Run

Prepare config.json and model.safetensors from the pinned Hugging Face revision in one directory, then run:

python inference.py --model-path ./weights --device npu --device-index 0 --output-dir ./validation/run

The temporary validation used LeRobot v0.3.3, whose Python requirement is compatible with the Pod Python 3.11 runtime. Dependencies are listed in requirements.txt.

Evidence

Agent workflow

NPU device call

Model result