What Is BridgeVLA?

BridgeVLA is introduced in BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models (arXiv 2506.07961; v1 9 Jun 2025, v2 14 Oct 2025) by Peiyan Li, Yixiang Chen, Hongtao Wu, Xiao Ma and five co-authors. The arXiv page lists NeurIPS 2025. Read: the arXiv abstract page and the GitHub README; the project page showed only a title and a continue prompt, and the full paper text was not read.

Third-party research; Firsthand is not affiliated and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: depth derived from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.

The abstract says pre-trained vision-language models are a promising base for VLA models, but few methods add 3D signals to them, and those do not fully use the spatial structure in 3D data, which leads to low sample efficiency.

Per the abstract, BridgeVLA does two things. It projects 3D inputs to multiple 2D images, so the input matches the VLM backbone, and it uses 2D heatmaps for action prediction, so inputs and outputs share one consistent 2D image space. The GitHub README describes the same flow: point clouds are rendered into multi-view images, and actions are predicted as heatmaps in that same space.

The abstract proposes a scalable pre-training method that teaches the VLM backbone to predict 2D heatmaps before downstream policy learning. The README adds that this pre-training predicts language-conditioned 2D heatmaps on object-detection data. The corpus size and the backbone are not given in the abstract.

The abstract reports that BridgeVLA outperforms baselines across three simulation benchmarks: RLBench average success rises from 81.4% to 88.2%, and COLOSSEUM from 56.7% to 64.0%. It also says BridgeVLA surpasses all compared baselines on GemBench. In real-robot tests it reports a 32% average gain over a baseline and 96.8% success on 10+ tasks with only 3 trajectories per task. These are the authors’ claims, not Firsthand results.

The site’s entries describe 3D-VLA as a generative world model that predicts goal images and point clouds, SpatialVLA as using Ego3D Position Encoding and Adaptive Action Grids, and PointVLA as adding point cloud features to a pre-trained VLA. BridgeVLA’s stated move is different again: render 3D into 2D images for the input and predict heatmaps for the output. The table ranks none of them.

BridgeVLA starts from 3D inputs. On the stereo capture pages, the site states depth derived from a head-mounted calibrated stereo rig, not a standalone depth sensor, in metric millimetres. The site’s point-cloud entry says a buyer can compute a cloud from the delivered depth and calibration; point-cloud file delivery is Not stated. Whether that suits BridgeVLA is the buyer’s to check.

The full paper was not read, so view count, heatmap resolution, how a heatmap becomes a robot action and the backbone are Not stated here. The README separates the original BridgeVLA code, on a bridgevla branch, from a later BridgeVLA++ on main; this page covers only BridgeVLA, and the README’s own results table gives different BridgeVLA figures from the abstract, so quote the paper for exact protocols. Real-robot data is described there as self-collected and not published.

Human video is not BridgeVLA’s stated input. Firsthand records first-person human video to a written spec: depth from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories, action segments and one task instruction per episode. That is the site’s only language data, and robot actions are Not stated. Confirm tasks and signals on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS; matching BridgeVLA’s expected fields is the buyer’s to check.

What is BridgeVLA?

A 3D vision-language-action model (arXiv 2506.07961) that projects 3D inputs to multiple 2D images and predicts actions as 2D heatmaps.

What does input-output alignment mean in BridgeVLA?

Per the abstract, the 3D input is projected to 2D images to match the VLM backbone, and actions are predicted as 2D heatmaps, so both sit in one consistent 2D image space.

How does BridgeVLA use heatmaps for action prediction?

The abstract says it uses 2D heatmaps for action prediction, after pre-training the VLM backbone to predict heatmaps. How a heatmap becomes a robot action is Not stated in the sources read.

How is BridgeVLA different from 3D-VLA, SpatialVLA and PointVLA?

It renders 3D input into 2D images and predicts heatmaps. The site entries describe 3D-VLA as a generative world model, SpatialVLA as using Ego3D Position Encoding and Adaptive Action Grids, and PointVLA as adding point cloud features to a pre-trained VLA.

Is BridgeVLA open source?

The GitHub page lists an official repository under Apache License 2.0. The README says the original BridgeVLA code is on a bridgevla branch and main hosts BridgeVLA++.

Does Firsthand use BridgeVLA, and can I request a custom dataset?

Firsthand is not affiliated and claims no use, training or compatibility. The site states depth from a calibrated stereo rig and one task instruction per episode. You can request a custom dataset, confirmed first on a Spec Pilot.