README.md

---
license: mit
tags:
  - vla
  - robotics
  - libero
  - qwen2.5
  - lora
  - vision-language-action
library_name: transformers
pipeline_tag: robotics
---

# Anchor-Align: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment

[**🌐 Project Page**](https://anchoralignvla.github.io/)  •  [**📄 Paper (arXiv)**](https://arxiv.org/abs/2607.13429)  •  [**💻 Code (GitHub)**](https://github.com/dwipddalal/Anchor-Align)  •  [**🎬 Videos**](https://anchoralignvla.github.io/#demo)

<p align="center">
  <img src="assets/fig_method_overview.png" alt="Anchor-Align method overview" width="95%"/>
</p>

**TL;DR** — Standard behavior-cloning finetuning of a vision-language model on robot demos silently erases the pretrained VLM's semantics and decouples its language output from its actions. **Anchor-Align** adds two lightweight losses on top of action prediction that (i) *anchor* the trainable VLA to a frozen copy of the pretrained VLM (preserving vision-language reasoning) and (ii) *align* the pre-action hidden state with a discrete motion-direction label derived from the executed motion. The result: **22.6%** on LIBERO-PRO position swap (vs 2.3% for standard BC), **90.3%** LIBERO-Plus average (vs 85.1%), and near-doubled real-world success (**28.3% → 54.2%**) on a UFactory xArm7 setup.

## 📦 Companion GitHub Repository

**This repo hosts the weights. The inference and evaluation code lives at [github.com/dwipddalal/Anchor-Align](https://github.com/dwipddalal/Anchor-Align).**

| | HuggingFace (`Dwipz/Anchor-Align`) | GitHub (`dwipddalal/Anchor-Align`) |
|---|---|---|
| Weights + tokenizer + configs | ✅ | — |
| LIBERO / LIBERO-PRO / LIBERO-Plus eval scripts | — | ✅ |
| Eval SLURM templates + 3-seed campaign runners | — | ✅ |
| Paper results tables + figures | ✅ | ✅ |
| Per-checkpoint `MODEL_CARD.md` | ✅ | — |
| Reproduction recipe (`REPRODUCE.md` + verifier) | — | ✅ |

> **Note:** the training code is not part of the initial release; we plan to release it in a follow-up update. Each checkpoint's full training configuration is documented in its `MODEL_CARD.md`.

## Available Checkpoints

| Folder | Suite | KL weight | Steps |
|---|---|---:|---:|
| [`libero-spatial/`](./libero-spatial) | LIBERO Spatial | 0.10 | 10k |
| [`libero-object/`](./libero-object) | LIBERO Object | 0.15 | 2.5k |
| [`libero-goal/`](./libero-goal) | LIBERO Goal | 0.10 | 25k |
| [`libero-long/`](./libero-long) | LIBERO-10 (Long) | 0.15 | 45k |

All four finetune **Prismatic Qwen2.5-0.5B** (DINOv2 + SigLIP, 24 layers) with LoRA rank 64 and an L1-regression MLPResNet action head. Per-checkpoint metrics live in each subfolder's `MODEL_CARD.md`; reproduction targets and tolerances are in the GitHub repo's [`REPRODUCE.md`](https://github.com/dwipddalal/Anchor-Align/blob/main/REPRODUCE.md).

## Results

Tables below are reproduced from the paper. Reproduction targets and tolerances for the released checkpoints are in [`REPRODUCE.md`](https://github.com/dwipddalal/Anchor-Align/blob/main/REPRODUCE.md).

### Standard LIBERO suites

Success rates on the four standard (unperturbed) LIBERO suites. Anchor-Align achieves the highest success rate on every suite, surpassing methods with substantially larger backbones and large-scale robot-action pretraining.

| Method | Spatial | Object | Goal | Long |
|---|---:|---:|---:|---:|
| Diffusion Policy | 78.3 | 92.5 | 68.3 | 50.5 |
| π₀-FAST | 87.0 | 63.0 | 89.0 | 48.0 |
| SmolVLA-0.24B | 87.0 | 93.0 | 88.0 | 63.0 |
| SmolVLA-2.25B | 93.0 | 94.0 | 91.0 | 77.0 |
| OpenVLA-OFT | 94.3 | 95.2 | 91.7 | 86.5 |
| MolmoAct | 87.0 | 95.4 | 87.6 | 77.2 |
| π₀.₅-KI | 96.6 | 97.2 | 94.6 | 85.8 |
| VLA-0 | 93.6 | 96.0 | 95.6 | 87.6 |
| VLA-Adapter [Frozen] | 89.4 | 89.6 | 88.0 | 84.5 |
| VLA-Adapter (standard BC) | 96.0 | 99.8 | 96.0 | 89.0 |
| **Anchor-Align VLA (ours)** | **98.4** | **100.0** | **97.2** | **90.8** |

### Robustness and generalization — LIBERO-PRO and LIBERO-Plus

Success rates under perturbation on the LIBERO-Spatial suite (paper Table 1). **Bold** = best, <u>underline</u> = second best.

<table>
<thead>
<tr><th rowspan="2" align="left">Method</th><th colspan="4">LIBERO-PRO</th><th colspan="8">LIBERO-Plus</th></tr>
<tr><th>Lang. Reph.</th><th>Object Swap</th><th>Pos. Swap</th><th>Mean</th><th>Lang. Instr.</th><th>Bg. Text.</th><th>Robot Init</th><th>Cam. View</th><th>Obj. Layout</th><th>Light Cond.</th><th>Sensor Noise</th><th>Mean</th></tr>
</thead>
<tbody>
<tr><td align="left">Co-training + KI*</td><td>54.0</td><td>77.4</td><td>0.0</td><td>43.8</td><td>48.0</td><td>82.6</td><td>25.7</td><td>64.6</td><td>65.7</td><td>73.3</td><td>49.0</td><td>57.1</td></tr>
<tr><td align="left">MolmoAct</td><td>77.8</td><td>82.4</td><td>0.0</td><td>53.4</td><td>79.5</td><td>84.1</td><td>47.4</td><td>10.1</td><td>76.5</td><td>77.4</td><td>53.4</td><td>60.8</td></tr>
<tr><td align="left">OpenVLA-OFT</td><td>74.4</td><td><u>95.2</u></td><td>0.0</td><td>56.5</td><td>81.5</td><td><u>95.7</u></td><td>40.3</td><td><u>94.7</u></td><td>88.6</td><td><u>95.5</u></td><td>28.2</td><td>74.1</td></tr>
<tr><td align="left">VLA-Adapter [Frozen]</td><td>56.0</td><td>73.4</td><td>0.0</td><td>43.1</td><td>41.5</td><td>70.9</td><td>35.1</td><td>94.4</td><td>62.3</td><td>84.9</td><td>36.2</td><td>59.9</td></tr>
<tr><td align="left">VLA-Adapter (standard BC)</td><td><u>91.1</u></td><td>89.6</td><td><u>2.3</u></td><td><u>61.0</u></td><td><u>85.1</u></td><td>90.7</td><td><u>52.6</u></td><td>92.6</td><td><u>93.2</u></td><td>93.2</td><td><u>89.5</u></td><td><u>85.1</u></td></tr>
<tr><td align="left"><b>Anchor-Align VLA (ours)</b></td><td><b>97.0</b></td><td><b>96.2</b></td><td><b>22.6</b></td><td><b>71.9</b></td><td><b>87.2</b></td><td><b>99.6</b></td><td><b>59.1</b></td><td><b>96.3</b></td><td><b>97.4</b></td><td><b>99.0</b></td><td><b>96.9</b></td><td><b>90.3</b></td></tr>
</tbody>
</table>

\*Our implementation of knowledge insulation adapted to VLA-Adapter. Position swap is the hardest axis: MolmoAct and OpenVLA-OFT score 0%, standard BC reaches 2.3%, while Anchor-Align reaches 22.6%.

### Qualitative — generalization to semantic perturbations

<p align="center">
  <img src="assets/fig3_qualitative_swap.png" alt="Anchor-Align generalizes to semantic perturbations" width="95%"/>
</p>

Same task, different phrasing / different object identity / shuffled positions. Anchor-Align retains the VLM's semantic understanding of what the instruction refers to, while the baseline latches onto memorized appearance shortcuts.

### Per-suite robustness — Object, Goal, and Long

The same gains carry over to the remaining three LIBERO suites. Each radar plot compares Anchor-Align (orange) against the standard BC VLA-Adapter baseline (gray) across nine evaluation axes: two from LIBERO-PRO (Language Rephrase, Object Swap) and seven from LIBERO-Plus.

<p align="center">
  <img src="assets/fig_object_spider.png" alt="LIBERO Object suite radar plot" width="32%"/>
  <img src="assets/fig_goal_spider.png" alt="LIBERO Goal suite radar plot" width="32%"/>
  <img src="assets/fig_long_spider.png" alt="LIBERO Long suite radar plot" width="32%"/>
</p>

Largest gains: Robot Init State +18.6 on Object; Language Instruction +11.9 on Goal; Lighting Condition +20.8, Object Layout +18.6, and Camera Viewpoint +17.7 on Long.

### Long-horizon generalization — CALVIN ABC→D

Each rollout chains five language instructions; k/5 is the fraction of rollouts completing the first k, and Len is the average number of consecutively completed tasks.

| Method | 1/5 | 2/5 | 3/5 | 4/5 | 5/5 | Len |
|---|---:|---:|---:|---:|---:|---:|
| UniVLA | 95.5 | 85.8 | 75.4 | 66.9 | 56.5 | 3.8 |
| OpenVLA-OFT | 96.3 | 89.1 | 82.4 | 75.8 | 66.5 | 4.1 |
| OpenHelix | 97.1 | 91.4 | 82.8 | 72.6 | 64.1 | 4.1 |
| VLA-Adapter (standard BC) | <u>98.3</u> | <u>94.0</u> | <u>87.5</u> | <u>80.0</u> | <u>73.1</u> | <u>4.3</u> |
| **Anchor-Align VLA (ours)** | **99.1** | **95.8** | **90.6** | **84.7** | **77.9** | **4.5** |

### Multi-seed significance

Mean ± standard deviation over 5 training seeds on LIBERO-Spatial; the method gaps are far larger than seed-to-seed variability.

| Method | Spatial (Std) | Lang. Reph. | Object Swap | Pos. Swap | Plus |
|---|---:|---:|---:|---:|---:|
| VLA-Adapter (standard BC) | 93.3 ± 0.3 | 91.1 ± 0.4 | 90.1 ± 0.5 | 2.6 ± 0.7 | 85.3 ± 0.3 |
| **Anchor-Align VLA (ours)** | **97.9 ± 0.3** | **97.1 ± 0.5** | **96.1 ± 0.4** | **23.5 ± 0.2** | **90.5 ± 0.6** |

---

## Per-folder layout

Each subfolder is a self-contained inference bundle:

```
<benchmark>_<steps>/
  MODEL_CARD.md
  model.safetensors                        # merged base VLM weights (~2.5 GB)
  lora_adapter/adapter_model.safetensors   # separate LoRA weights (~208 MB)
  action_head--<step>_checkpoint.pt        # L1 regression head (~436 MB)
  align_dir_proj--<step>_checkpoint.pt     # alignment direction projector
  proprio_projector--<step>_checkpoint.pt  # proprioception projector
  config.json, tokenizer.*, processor_config.json, dataset_statistics.json
```

## Quickstart — download + run inference

**1. Clone the code repo** (contains all inference and eval scripts):

```bash
git clone https://github.com/dwipddalal/Anchor-Align.git
cd Anchor-Align
pip install -e .   # + follow the installation section in the GitHub README
```

**2. Download a checkpoint from this repo:**

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="Dwipz/Anchor-Align",
    allow_patterns="libero-spatial/*",   # or libero-object/*, libero-goal/*, libero-long/*
    local_dir="./checkpoints",
)
```

**3. Run inference** using the eval scripts in the GitHub repo:

```bash
# LIBERO Standard
CUDA_VISIBLE_DEVICES=0 python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint ./checkpoints/libero-spatial \
  --task_suite_name libero_spatial \
  --use_proprio True --num_images_in_input 2 --use_pro_version True

# LIBERO-PRO (paraphrased-language perturbation)
CUDA_VISIBLE_DEVICES=0 python experiments/robot/libero_pro/run_libero_pro_eval.py \
  --pretrained_checkpoint ./checkpoints/libero-spatial \
  --base_suite_name libero_spatial --perturbation_type lan \
  --use_proprio True --num_images_in_input 2 --use_pro_version True
```

Inference flags (`--use_proprio True --num_images_in_input 2 --use_pro_version True --use_l1_regression True --center_crop True --num_open_loop_steps 8`) match the paper. See the GitHub README's "Evaluation" section for LIBERO-Plus and batched variants.

## License

MIT. See the [GitHub repository LICENSE](https://github.com/dwipddalal/Anchor-Align/blob/main/LICENSE).

## Citation

```bibtex
@article{dalal2026anchoralign,
  title   = {Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment},
  author  = {Dalal, Dwip and Patel, Shivansh and Jain, Chahit and Kim, Jeonghwan and Mishra, Utkarsh and Baratian, Alex and Ha, Hyeonjeong and Ji, Heng and Lazebnik, Svetlana and Jain, Unnat},
  journal = {arXiv preprint arXiv:2607.13429},
  year    = {2026}
}
```