Update README.md
#2
by kgubernatorov - opened
README.md
CHANGED
|
@@ -10,13 +10,13 @@ tags:
|
|
| 10 |
base_model: lerobot/smolvla_base
|
| 11 |
---
|
| 12 |
|
| 13 |
-
#
|
| 14 |
|
| 15 |
Distilled from a tactile-conditioned [SA-RWFM teacher](https://huggingface.co/armteam/crab-smolvla-rwfm) via offline action-level knowledge distillation. At inference, this model is an **unmodified SmolVLA** β no tactile sensors, no extra modules, zero overhead.
|
| 16 |
|
| 17 |
## Key Result
|
| 18 |
|
| 19 |
-
|
| 20 |
|
| 21 |
## Model Details
|
| 22 |
|
|
@@ -42,7 +42,7 @@ HapticsVLA achieves **86.7% mean success rate** (vs. 61.7% baseline, 75.0% tacti
|
|
| 42 |
|-------|-------------|-----------------|-----------------|
|
| 43 |
| SmolVLA (Baseline) | 61.7% | 26.7% | No |
|
| 44 |
| SA-RWFM (Teacher) | 75.0% | 11.7% | **Yes** |
|
| 45 |
-
| **
|
| 46 |
|
| 47 |
## Distillation Approach
|
| 48 |
|
|
@@ -67,13 +67,13 @@ See [Advanced-Robotic-Manipulation/crab](https://github.com/Advanced-Robotic-Man
|
|
| 67 |
- [crab-smolvla-rwfm](https://huggingface.co/armteam/crab-smolvla-rwfm) β SA-RWFM tactile teacher
|
| 68 |
|
| 69 |
|
| 70 |
-
##
|
| 71 |
|
| 72 |
If you use this model, please cite our paper:
|
| 73 |
|
| 74 |
```bibtex
|
| 75 |
-
@article{
|
| 76 |
-
title={
|
| 77 |
author={Gubernatorov, Konstantin and Sannikov, Mikhail and Mikhalchuk, Ilya and Kuznetsov, Egor and Artemov, Makar and Ouwatobi, Ogunwoye Faith and Fernando, Marcelino and Asanov, Artem and Guo, Ziang and Tsetserukou, Dzmitry},
|
| 78 |
journal={arXiv preprint arXiv:2603.15257},
|
| 79 |
year={2026}
|
|
|
|
| 10 |
base_model: lerobot/smolvla_base
|
| 11 |
---
|
| 12 |
|
| 13 |
+
# HapticVLA β Contact-Rich Manipulation without Inference-Time Tactile Sensing
|
| 14 |
|
| 15 |
Distilled from a tactile-conditioned [SA-RWFM teacher](https://huggingface.co/armteam/crab-smolvla-rwfm) via offline action-level knowledge distillation. At inference, this model is an **unmodified SmolVLA** β no tactile sensors, no extra modules, zero overhead.
|
| 16 |
|
| 17 |
## Key Result
|
| 18 |
|
| 19 |
+
HapticVLA achieves **86.7% mean success rate** (vs. 61.7% baseline, 75.0% tactile teacher) with the **lowest force error rate of any model (5.0%)**, including the tactile-equipped teacher β all without requiring tactile sensors at inference.
|
| 20 |
|
| 21 |
## Model Details
|
| 22 |
|
|
|
|
| 42 |
|-------|-------------|-----------------|-----------------|
|
| 43 |
| SmolVLA (Baseline) | 61.7% | 26.7% | No |
|
| 44 |
| SA-RWFM (Teacher) | 75.0% | 11.7% | **Yes** |
|
| 45 |
+
| **HapticVLA (Ours)** | **86.7%** | **5.0%** | **No** |
|
| 46 |
|
| 47 |
## Distillation Approach
|
| 48 |
|
|
|
|
| 67 |
- [crab-smolvla-rwfm](https://huggingface.co/armteam/crab-smolvla-rwfm) β SA-RWFM tactile teacher
|
| 68 |
|
| 69 |
|
| 70 |
+
## Citation
|
| 71 |
|
| 72 |
If you use this model, please cite our paper:
|
| 73 |
|
| 74 |
```bibtex
|
| 75 |
+
@article{gubernatorov2026hapticvla,
|
| 76 |
+
title={HapticVLA: Contact-Rich Manipulation via Vision-Language-Action Model without Inference-Time Tactile Sensing},
|
| 77 |
author={Gubernatorov, Konstantin and Sannikov, Mikhail and Mikhalchuk, Ilya and Kuznetsov, Egor and Artemov, Makar and Ouwatobi, Ogunwoye Faith and Fernando, Marcelino and Asanov, Artem and Guo, Ziang and Tsetserukou, Dzmitry},
|
| 78 |
journal={arXiv preprint arXiv:2603.15257},
|
| 79 |
year={2026}
|