---
license: cc-by-4.0
pretty_name: LADDER offline VLM benchmark
task_categories:
- robotics
- visual-question-answering
tags:
- robotics
- manipulation
- long-horizon
- vlm
- benchmark
language:
- en
configs:
- config_name: decomposition
  data_files:
  - split: train
    path: decomposition/train-*
- config_name: subtask_states
  data_files:
  - split: train
    path: subtask_states/train-*
---

# LADDER offline VLM benchmark

[📄 Paper](https://arxiv.org/abs/2610.18053) · [💻 Code](https://github.com/insait-institute/ladder) · [🌐 Website](https://ladderrobotics.ai) · [🤗 LADDER collection](https://huggingface.co/collections/INSAIT-Institute/ladder-6abbbddc46691080f5d954c8)

The offline benchmark of [LADDER](https://github.com/insait-institute/ladder), a benchmark for long-horizon robot manipulation. It
holds the judgements an orchestrator asks a vision-language model (VLM) to make, on frames of the
three LADDER suites, scored without running a simulator. The labels come from the simulator's own
success checks.

| config | rows | input | tasks |
|---|---|---|---|
| `decomposition` | 145 | the first frame and the instruction of every task | plan decomposition |
| `subtask_states` | 562 | a frame from a recorded episode, with the true state of every subgoal | subgoal verification, state tracking, subgoal switching |

## Tasks

| task | config | the model gets | it answers |
|---|---|---|---|
| plan decomposition | `decomposition` | frame and instruction | the list of subgoals |
| subgoal verification | `subtask_states` | frame and one subgoal | done or not done |
| state tracking | `subtask_states` | frame and the plan | which subgoals are done |
| subgoal switching | `subtask_states` | frame and instruction | the next subgoal |

## Columns

`decomposition`: `id`, `benchmark` (the suite), `scene`, `instruction`, `subtasks` (the reference
plan), `n_goals`, `ordered`, `revisits`, `family`, `distractors`, `objects`, `image_main`, and on
RoboLab-Long `image_wrist`.

`subtask_states`: the same task columns, plus `achieved` (the state of every subgoal at this frame),
`query_idx` and `label` (the subgoal asked about in verification, and its answer), `tags` (what the
frame shows: `fresh`, `in_progress`, `regressed`, `cold`, `terminal`, `held`), `step`, `num_steps`,
`arm` and `source` (the episode it was cut from), and the images.

## Usage

Load a config with `datasets`:

```python
from datasets import load_dataset
plans = load_dataset("INSAIT-Institute/LADDER-Offline-VLM-Benchmark", "decomposition", split="train")
states = load_dataset("INSAIT-Institute/LADDER-Offline-VLM-Benchmark", "subtask_states", split="train")
```

Score a model with the [LADDER code](https://github.com/insait-institute/ladder), on an OpenAI-compatible endpoint
(`--backend gateway`, with `VLM_BASE_URL` / `VLM_API_KEY`) or locally from Hugging Face:

```bash
scripts/vlm_benchmark.sh decomposition --model <model>
scripts/vlm_benchmark.sh verification --model <model>
```

The first argument is `decomposition`, `verification`, `state_tracking` or `next_subtask`. The scorers use the orchestrator's own
components and prompts, so a score here is the model's score inside the orchestrator. The metrics
are described in
[`api/client/metrics/README.md`](https://github.com/insait-institute/ladder/blob/main/ladder/components/inference/api/client/metrics/README.md).

## Results

From the paper, in %. ± is the 95% Wilson interval, widened for the correlation of frames from the
same scene. TP and TN rates measure detecting finished and unfinished subgoals. Tokens are those
generated per query, reasoning included.

| model | planning: exact match | subgoal completion: TP rate | subgoal completion: TN rate | state tracking: all correct | next subgoal: valid | tokens per query |
|---|---:|---:|---:|---:|---:|---:|
| Qwen3-VL-2B | 75.2 ± 7.6 | 68.2 ± 7.0 | 34.1 ± 5.9 | 9.3 ± 4.0 | 58.9 ± 5.3 |  |
| Qwen3-VL-8B | 91.0 ± 5.8 | 69.0 ± 6.8 | 79.9 ± 5.7 | 47.9 ± 5.5 | 72.3 ± 4.4 |  |
| Qwen3.5-9B | 86.9 ± 6.5 | 44.4 ± 6.3 | 96.9 ± 2.5 | 11.7 ± 3.7 | 63.2 ± 4.8 | 5,173 |
| Qwen3.8-27B | 91.0 ± 5.8 | **76.2 ± 6.3** | 84.8 ± 4.3 | 58.2 ± 4.8 | 79.3 ± 4.1 | 1,904 |
| Gemini 3.1 Pro | 89.7 ± 6.0 | 54.8 ± 6.6 | 96.0 ± 2.9 | 68.7 ± 4.3 | 75.2 ± 4.5 | 583 |
| Gemini 3.7 Flash | **93.1 ± 5.4** | 59.4 ± 6.6 | 96.6 ± 2.6 | **70.8 ± 4.0** | 81.4 ± 4.2 | 412 |
| Gemini Robotics-ER 2 | 90.3 ± 5.9 | 54.0 ± 6.5 | **97.8 ± 2.3** | 69.6 ± 4.2 | 76.6 ± 4.1 | 510 |
| GPT-5.6 Sol | **93.1 ± 5.4** | 72.8 ± 6.3 | 86.4 ± 4.3 | 56.4 ± 4.7 | 82.0 ± 4.3 | 282 |
| GPT-6 Astra | 91.7 ± 5.7 | 71.1 ± 6.2 | 86.4 ± 4.4 | 53.2 ± 5.1 | **85.4 ± 4.0** | 203 |
| Claude Opus 5 | 90.3 ± 5.9 | 68.6 ± 6.6 | 87.9 ± 4.1 | 61.0 ± 4.7 | 79.8 ± 4.4 | 400 |

## License

CC BY 4.0. The RoboLab-Long frames show [RoboLab](https://github.com/NVlabs/RoboLab) assets, which are licensed under CC BY-NC-SA 4.0.

## Citation

```bibtex
@article{ladder2026,
  title   = {{LADDER}: A Long-Horizon Robot Manipulation Benchmark},
  author  = {Petkov, Petko and Albanese, Giuliano and Nikolov, Nikolay and Van Gool, Luc and Tremblay, Jonathan and Paudel, Danda Pani},
  journal = {arXiv preprint arXiv:2610.18053},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.18053}
}
```