README.md

---
license: mit
library_name: pointmapvla
pipeline_tag: robotics
tags:
  - robotics
  - vision-language-action
  - pointmap
  - qwen3.5
  - flow-matching
  - libero
---

# PointmapVLA-LIBERO

PointmapVLA fine-tuned on the four LIBERO suites (spatial, object, goal, 10). This is the
checkpoint behind the LIBERO result of the paper and, evaluated zero-shot, the LIBERO-Plus
result. Code, recipes and the evaluation harness:
[PointmapVLA repository](https://github.com/DAVIAN-Robotics/PointmapVLA).

## What it is

PointmapVLA is a vision-language-action model built on the Qwen3.5-4B vision-language
backbone with a **pointmap encoder** (a copy of the vision tower fed raw metric XYZ
pointmaps in the robot-base frame, features added patch-wise to the RGB tokens) and a
**32-layer layer-wise flow-matching DiT action expert** (4 Euler steps at inference).
6.08 B parameters, bf16.

This checkpoint predicts chunks of 8 LIBERO actions (10-D) from two camera views
(agentview, wrist), the instruction and an 11-D end-effector state.

## Provenance

* Initialised from [`DAVIAN-Robotics/PointmapVLA-Pretrained`](https://huggingface.co/DAVIAN-Robotics/PointmapVLA-Pretrained)
  (Stage 1 + Stage 2 pretraining on 76.3 M robot frames and 10.7 M 3D spatial VQA samples).
* Fine-tuned for 30,000 optimizer steps, global batch 128 (4 GPUs x 32), AdamW, cosine
  schedule with 1,500 warm-up steps, learning rates 1e-5 (backbone, pointmap encoder) and
  1e-4 (action expert, projections), DeepSpeed ZeRO-2, bf16. One policy for all four suites.
* Data: [`DAVIAN-Robotics/libero-r6`](https://huggingface.co/datasets/DAVIAN-Robotics/libero-r6)
  + [`DAVIAN-Robotics/libero-r6-pointmap`](https://huggingface.co/datasets/DAVIAN-Robotics/libero-r6-pointmap)
  (1,693 episodes over 40 tasks, 2 cameras).
* On load from the pretrained checkpoint the four action-head tensors whose shape depends
  on the layout were remapped (shared end-effector/gripper slots copied, new slots fresh);
  all other weights transferred unchanged.
* Prompt: `Task: {instruction} Mode: <end_effector> Pointmap_origin: <robot>`, chat
  template default for thinking (no `enable_thinking` key; differs from the RoboCasa
  checkpoint and is recorded in `config.yaml`).
* Exact config: `examples/simBenchmarks/LIBERO/train_files/pointmapvla_libero.yaml` in
  the repository (`_80gb.yaml` is the same recipe with gradient accumulation).

## Files

```
config.yaml                        architecture and the data recipe (release schema)
dataset_statistics.json            action/state mean/std/min/max/quantiles used for (un)normalisation
base_vlm/                          Qwen3.5-4B config, tokenizer and processor only; no weights
checkpoints/
  model-00001-of-00003.safetensors
  model-00002-of-00003.safetensors
  model-00003-of-00003.safetensors   ~11 GB in total, bf16
  model.safetensors.index.json
```

State-dict key groups: `qwen_vl_interface.model.*` (backbone), `pointmap_encoder.*`,
`action_model.*` (DiT expert), `project_layers.*` (per-layer projections). The language
model head is tied to the token embedding and stored once; the index records it under
`metadata.tied_weights` and the loader re-ties it.

## How to load

```python
from starVLA.model.framework.VLM4A.PointmapVLA import PointmapVLA

model = PointmapVLA.from_pretrained("DAVIAN-Robotics/PointmapVLA-LIBERO")
out = model.predict_action(examples)                    # {"normalized_actions": (B, 8, 10)}
```

`examples[i]` is a dict with `image` (2 HWC uint8 arrays, 224 x 224, agentview and wrist,
rotated 180° as in the LIBERO datasets), `lang`, `state` (`(1, 11)` float32, normalised),
`pointmap` (`(2, 256, 256, 3)` float16 metres, robot-base frame) and `pointmap_valid`
(`(2, 256, 256)` bool). Un-normalise with `dataset_statistics.json`. In practice, run the
policy server from the repository, which does the un-normalisation and the r6 ->
axis-angle conversion:

```bash
python deployment/model_server/server_policy.py --ckpt DAVIAN-Robotics/PointmapVLA-LIBERO --port 21000 --use_bf16 --seed 7
```

## Evaluation

Protocol: action chunk 8 executed in full, 4 Euler steps, K = 12 parallel environments,
server + environment seeds 7, 8, 9 (one run each); numbers are the mean (± sd) over the
three seeds. LIBERO: 4 suites x 10 tasks x 50 trials = 2,000 episodes. LIBERO-Plus
(zero-shot, same checkpoint): 10,030 perturbed tasks x 1 trial (2402 / 2518 / 2591 / 2519
per suite). One seed alone typically lands within about ±0.2 points of the LIBERO mean.

```bash
for s in 7 8 9; do CKPT=DAVIAN-Robotics/PointmapVLA-LIBERO BENCH=libero      SEED=$s scripts/eval/run_eval_libero.sh; done
for s in 7 8 9; do CKPT=DAVIAN-Robotics/PointmapVLA-LIBERO BENCH=libero_plus SEED=$s GPUS=0,1 scripts/eval/run_eval_libero.sh; done
```

| Benchmark | Paper | Our reproduction |
|---|---|---|
| LIBERO (4 suites), success rate (%) | 99.0 | 99.1 ± 0.1 (seeds 7, 8, 9) |
| LIBERO-Plus, success rate (%) | 92.4 | 92.4 ± 0.3 (seeds 7, 8, 9) |

## Intended use and limitations

* Simulation benchmark checkpoint for LIBERO and LIBERO-Plus (Franka Panda). Not
  validated on physical robots.
* Layouts are fixed: action 10-D (EE position 3 | EE rotation r6 6 | gripper 1), horizon
  8; state 11-D (EE position 3 | EE rotation r6 6 | gripper 2). For another robot or
  layout, fine-tune `PointmapVLA-Pretrained` instead.
* Requires metric depth with known intrinsics and camera-to-world extrinsics for the
  pointmap input; the model was trained with pointmaps present. LIBERO images are
  rotated by 180°, encoded in the datasets as a negative `fx`.

## License

MIT (code and weights of this release). Weights are derived from Qwen3.5-4B, distributed
under the Apache License 2.0. See `NOTICE` in the repository.

## Citation

TBD. The paper is under double-blind review; the BibTeX entry and author list will be
added at camera-ready.