---
license: apache-2.0
tags:
  - robotics
  - vision-language-action
  - manipulation
---

# BridgeVLA++

Pre-training data, checkpoints and benchmark keyframe data for **BridgeVLA++** and its
predecessor **BridgeVLA**.

BridgeVLA++ is a 3D vision-language-action framework that preserves the input-output
alignment of a pre-trained VLM during 3D action learning — point clouds are projected
into multi-view images and intermediate heatmaps are predicted before actions — and
extends it with a unified **spatio-temporal memory** modeling persistent spatial
context and temporal interaction history. It matches or surpasses BridgeVLA on the
original benchmarks without sacrificing data efficiency or generalization, reaches
state of the art on two memory-dependent benchmarks, and extends to bimanual
manipulation and a new real-world embodiment.

## Contents

```
checkpoints/
├── pretrain/              # grounding pre-training weights, shared by both models (finetune warm-start)
├── bridgevla/             # BridgeVLA (original)
│   └── rlbench/ colosseum/ gembench/
└── bridgevla_plus/        # BridgeVLA++ weights
    ├── rlbench/ colosseum/ gembench/ memorybench/
    └── rmbench/<task>/    # per-task, 9 tasks
pretrain_data/
├── coco.tar.gz            # COCO images
└── detection_data.json    # RoboPoint grounding annotations
datasets/                  # benchmark keyframe data (see below)
├── rlbench/keyframe_cache/size128_v2/<task>/episode<N>.npz + .npz.meta
├── memorybench/keyframe_cache/size128_v3/<task>/episode<N>.npz + .npz.meta
└── rmbench/
    ├── keyframe_data/<task>/keyframe_depth/   # keyframe-only HDF5 training data
    └── keyframes/<task>.json                  # keyframe metadata -> memory labels
```

### Benchmark keyframe data

`datasets/` ships the precomputed keyframe artifacts training depends on — do not
rearrange them by hand; the code repo's `scripts/download_checkpoints_hf.sh` (or
`scripts/download_checkpoints_ms.sh` for the identical ModelScope mirror) with a
target of `rlbench_cache` / `memorybench_cache` / `rmbench_data` (or per-task
`rmbench_data:<task>`) places each into the exact layout the trainers expect:

- **rlbench / memorybench `keyframe_cache`** — pre-built episode caches
  (`.npz` decoded observations + `.meta` canonical keyframe indices). The
  `.meta` files pin RLBench's per-(task, variation) majority-vote canonical
  keyframes: shipping them makes training runs reproduce ours exactly, and
  skips a multi-hour local build. RLBench training *requires* this cache.
- **rmbench `keyframe_data` + `keyframes`** — the actual RMBench training set
  (keyframe-only re-rendered HDF5, 50 episodes x 10 tasks) plus per-keyframe
  metadata whose `language_annotation`/`subtask_idx` provide the memory
  supervision labels. With these, the raw 37 GiB demo_clean demos are NOT
  needed for training or evaluation. The two directories must stay siblings.

If you download this repo manually instead (e.g. `huggingface-cli download` /
`modelscope download` of the whole repo), place each tree as below. The local
names differ from the repo paths **on purpose** (`_keyframe_cache` has a leading
underscore; capitalization and nesting differ too), so copying `datasets/` into
the code repo as-is will NOT be found by the trainers:

| in this repo | local path in the code repo |
|---|---|
| `checkpoints/` | `data/bridgevla_ckpt/` (inner layout unchanged) |
| `pretrain_data/` | `data/bridgevla_data/pretrain_data/` |
| `datasets/rlbench/keyframe_cache/` | `data/bridgevla_data/RLBench/_keyframe_cache/` |
| `datasets/memorybench/keyframe_cache/` | `data/bridgevla_data/memorybench/data/train/_keyframe_cache/` |
| `datasets/rmbench/keyframe_data/` | `data/bridgevla_data/RMBench/data/keyframe_data/` |
| `datasets/rmbench/keyframes/` | `data/bridgevla_data/RMBench/data/keyframes/` |

A misplaced path never degrades silently: training fails fast with an error
naming the expected location and the download command that fills it.

Every checkpoint directory holds `model_<epoch>.pth` together with `exp_cfg.yaml` and
`mvt_cfg.yaml`; the two configs define the network architecture and must stay next to
the weights.

| epoch | RLBench | COLOSSEUM | GemBench | memoryBench | RMBench |
|---|---|---|---|---|---|
| BridgeVLA++ | 130 | 200 | 200 | 160 | per-task |
| BridgeVLA | 80 | 80 | 40 | – | – |

The `bridgevla/` checkpoints belong to the original BridgeVLA codebase and are **not**
loadable by BridgeVLA++; run them with that codebase. Only `pretrain/` is shared by both.

## Papers

- **BridgeVLA++: A Data-Efficient, Generalizable, and Memory-Augmented Vision-Language-Action Framework for 3D Manipulation** — arXiv coming soon.
- **BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models** — [arXiv:2506.07961](https://arxiv.org/abs/2506.07961)

Usage instructions live in the code repository.

## Provenance

`pretrain/`, `bridgevla/` and `pretrain_data/` are the artifacts released with
BridgeVLA. `pretrain_data/` builds on COCO images and RoboPoint-style grounding
annotations, which remain subject to their original terms; the Apache-2.0 license above
applies to the model weights.

`datasets/` is derived data: the rlbench cache from the PerAct RLBench demos
(hqfang/rlbench-18-tasks), the memorybench cache from SAM2Act's MemoryBench data
(hqfang/memorybench), and the rmbench trees re-rendered from RoboTwin 2.0 / RMBench
(TianxingChen/RMBench). Each remains subject to its upstream benchmark's terms.