凤
weixin_44093752/anikitakis_vla_so101_pick_n_place_full_expert_Ascend_NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

anikitakis/vla_so101_pick_n_place_full_expert Ascend NPU adaptation

This delivery keeps the official LeRobot SmolVLA checkpoint and executes its official policy on Ascend910 npu:0. The model accepts three RGB camera views with shape 3x480x640, a 6-dimensional robot state, and a natural-language task instruction. It returns a 50-step action chunk with 6 values per step.

Official source

  • Hugging Face model: https://huggingface.co/anikitakis/vla_so101_pick_n_place_full_expert
  • Revision: adda798b270742f820834daea0fc37718745f025
  • Checkpoint: model.safetensors, 906,712,520 bytes
  • SHA-256: 6cfe9930ccd0107242a74b6458de611d644c46a73a4e02c4a108a1bc8e41974c
  • Official runtime: https://github.com/huggingface/lerobot/tree/v0.3.3
  • Base VLM: HuggingFaceTB/SmolVLM2-500M-Video-Instruct

NPU smoke result

The official 500-key checkpoint loaded into the official SmolVLA policy with zero load errors. The policy has 450,046,212 parameters. On a real Ascend910 npu:0, the three-camera and 6-state input contract completed in 0.5953960530 seconds in the delivery-script run. The output was float32, shape [1, 50, 6], was produced on npu:0, and all values were finite. This is an NPU smoke gate only; full robotics-task accuracy was not evaluated.

Run

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

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

Dependencies are listed in requirements.txt. The temporary Ascend validation used LeRobot v0.3.3 and placed the Hugging Face cache under /tmp.

Evidence

Agent workflow

NPU device call

Model result