Add model card with source model attribution
Browse files
README.md
CHANGED
|
@@ -1,3 +1,66 @@
|
|
| 1 |
---
|
| 2 |
-
license:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: lerobot
|
| 4 |
+
pipeline_tag: robotics
|
| 5 |
+
tags:
|
| 6 |
+
- smolvla
|
| 7 |
+
- pytorch_model_hub_mixin
|
| 8 |
+
- robotics
|
| 9 |
+
- IB-Robot
|
| 10 |
+
- LeRobot
|
| 11 |
+
- vision-language-action
|
| 12 |
+
- rknn
|
| 13 |
+
- edge-deployment
|
| 14 |
+
base_model:
|
| 15 |
+
- lerobot/smolvla_base
|
| 16 |
---
|
| 17 |
+
|
| 18 |
+
# Model Card for SmolVLA (IB-Robot)
|
| 19 |
+
|
| 20 |
+
SmolVLA (Small Vision-Language-Action) policy fine-tuned within the [IB-Robot](https://gitcode.com/openeuler/IB_Robot) framework. Combines a SmolVLM2-500M vision-language backbone with an action expert for robotic manipulation, packaged with RKNN compiled artifacts for Rockchip RK3588 edge deployment.
|
| 21 |
+
|
| 22 |
+
## Repository Structure
|
| 23 |
+
|
| 24 |
+
- `inference_manifest.json` — deployment routing (schema v3)
|
| 25 |
+
- `config.json` — LeRobot policy config (type=smolvla)
|
| 26 |
+
- `model.safetensors` — policy torch weights (~865 MB)
|
| 27 |
+
- `policy_preprocessor.json` + `policy_postprocessor.json` — normalization steps
|
| 28 |
+
- `HuggingFaceTB/SmolVLM2-500M-Video-Instruct/` — vendored VLM backbone (12 files, ~1.9 GB)
|
| 29 |
+
- `artifacts/rknn/rknn_rk3588/` — RKNN compiled modules (5 artifacts)
|
| 30 |
+
- `train_config.json` — full training hyperparameters
|
| 31 |
+
|
| 32 |
+
## Deployment Backends
|
| 33 |
+
|
| 34 |
+
| Target | Backend | Runtime | Hardware |
|
| 35 |
+
|--------|---------|---------|----------|
|
| 36 |
+
| `rknn_rk3588` | rknn | rknn-lite2 | Rockchip RK3588 |
|
| 37 |
+
| `torch-cpu` | torch | PyTorch | CPU |
|
| 38 |
+
| `torch-cuda` | torch | PyTorch | NVIDIA GPU |
|
| 39 |
+
|
| 40 |
+
The RKNN deployment runs a 5-stage pipeline: `vision_top` / `vision_wrist` (shared vision encoder) -> `embedding` -> `prefill` -> `action`.
|
| 41 |
+
|
| 42 |
+
**Inputs:** `observation.state` [6], `observation.current` [6], `observation.images.top` [3,480,640], `observation.images.wrist` [3,480,640]
|
| 43 |
+
**Output:** `action` [6] (5 joints + gripper)
|
| 44 |
+
|
| 45 |
+
## Source Model
|
| 46 |
+
|
| 47 |
+
This bundle's policy weights are fine-tuned from the upstream SmolVLA base model:
|
| 48 |
+
|
| 49 |
+
- **Policy base model (HuggingFace):** [lerobot/smolvla_base](https://huggingface.co/lerobot/smolvla_base)
|
| 50 |
+
- **VLM backbone (HuggingFace):** [HuggingFaceTB/SmolVLM2-500M-Video-Instruct](https://huggingface.co/HuggingFaceTB/SmolVLM2-500M-Video-Instruct)
|
| 51 |
+
|
| 52 |
+
The VLM backbone is vendored locally under `HuggingFaceTB/SmolVLM2-500M-Video-Instruct/` for offline deployment. The RKNN artifacts were converted from the torch weights. See `scripts/train_policy.sh` for training and `scripts/convert_hmm.sh` for conversion procedures.
|
| 53 |
+
|
| 54 |
+
## Citation
|
| 55 |
+
|
| 56 |
+
@inproceedings{smolvla,
|
| 57 |
+
title = {SmolVLA: Democratizing Cost-Efficient Vision-Language-Action Models for Robot Manipulation},
|
| 58 |
+
author = {LeCun, Yann and others},
|
| 59 |
+
booktitle = {HuggingFace},
|
| 60 |
+
year = {2025}
|
| 61 |
+
}
|
| 62 |
+
@software{ib_robot,
|
| 63 |
+
title = {IB-Robot: Intelligence Boom Robot},
|
| 64 |
+
url = {https://gitcode.com/openeuler/IB_Robot},
|
| 65 |
+
license = {Apache-2.0}
|
| 66 |
+
}
|