What Is Long-VLA?

Long-VLA is introduced in Long-VLA: Unleashing Long-Horizon Capability of Vision Language Action Model for Robot Manipulation (arXiv 2508.19958; submitted 27 Aug 2025, cs.RO; accepted to CoRL 2025 per the arXiv comments) by Yiguo Fan, Pengxiang Ding, Shuanghao Bai, Xinyang Tong, Yuyang Zhu and eight co-authors. Read: the arXiv abstract page, the official project page and the GitHub repository long-vla/long-vla.github.io, which holds the project page source; the retrieved repository page showed no README text or code, and no separate code repository was found. 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: 21 keypoints per hand with a visibility flag, wrist trajectories and ordered action segments, plus one task instruction per episode.

The abstract says VLA models have become a cornerstone of robotic policy learning, but that existing VLA frameworks primarily address short-horizon tasks. Their effectiveness on long-horizon, multi-step manipulation remains limited, the authors write, because of challenges in skill chaining and subtask dependencies. Long-VLA is presented as the first end-to-end VLA model specifically designed for long-horizon robotic tasks; that is the authors’ claim, not a Firsthand finding. Its central feature is a phase-aware input masking strategy. That is what the phrases Long-VLA long-horizon VLA and long-horizon robot manipulation vision-language-action refer to in the paper.

Per the abstract, the strategy adaptively segments each subtask into moving and interaction phases, so the model can focus on phase-relevant sensory cues and the subtasks become more compatible with each other. The project page describes a three-part framework. First, data and phase decomposition: task decomposition with aligned visual observations and language annotations. Second, input-level adaptation via masking: phase-aware masking lets the model selectively attend to relevant tokens during attention computation without modifying the input structure. Third, end-to-end training on the decomposed data with phase-aware masking. The abstract adds that the strategy preserves the scalability and data efficiency of VLA training, and that the module is architecture-agnostic, so it can be integrated into existing VLA models.

The abstract says the authors propose the L-CALVIN benchmark to evaluate long-horizon manipulation systematically, and that experiments on simulated and real-world tasks show Long-VLA significantly outperforming prior state-of-the-art methods. The project page names two real-world task families, Sorting and Cleaning, and says Long-VLA consistently outperformed both the Base Policy and the current state-of-the-art method, pi0. It also says performance stayed robust under unseen lighting conditions and visual distraction, and shows demos such as cleaning up the kitchen and sorting cubes in a given order. These are the authors’ claims, not Firsthand results, and this page repeats no scores because none were read.

The site’s long-horizon task entry defines a kind of job: several sub-actions completed in order toward one outcome. Long-VLA is a model built for such jobs, not the job itself. The site’s MemoryVLA entry describes a separate published framework that adds a perceptual-cognitive memory bank for temporal context on long-horizon manipulation. The two papers are different, and the sources read here do not compare them. The site’s task decomposition entry describes breaking a human job into short, repeatable task units. Long-VLA’s phase decomposition is a different cut: it splits each subtask into moving and interaction phases. The table ranks none of them.

The full paper was not read, so the base policy, the model backbone, how the moving and interaction phases are detected or labelled, the exact masking rules, how L-CALVIN is constructed, training settings, dataset sizes and per-task numbers are Not stated here. The project page text gives no scores. The GitHub repository found holds the project page source only, and no code or checkpoint release was found, so availability and licence are Not stated. Quote the paper for exact protocols.

The project page says training uses decomposed data with aligned visual observations and language annotations. Firsthand’s site states human video with depth derived from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories, and ordered, timestamped action segments within each episode. The site’s only language data is one task instruction per episode. Whether Firsthand’s action segments would map onto moving and interaction phases is Not stated, and phase labels are not a stated deliverable. Robot actions are not delivered either.

The sources read describe a robot model trained on decomposed, language-annotated data, not human video. 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, ordered action segments and one task instruction per episode. Whether such data suits a given long-horizon VLA is the buyer’s to check; the sources do not say. Confirm tasks and signals on a Spec Pilot, and use fast, on-demand custom collection when a long task needs its own spec. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS, LeRobot format and WebDataset.

What is Long-VLA?

A published vision-language-action model (arXiv 2508.19958) for long-horizon manipulation. Per the abstract it uses phase-aware input masking and an architecture-agnostic module that can be integrated into existing VLAs.

What is phase-aware masking in Long-VLA?

Per the abstract, a strategy that segments each subtask into moving and interaction phases so the model focuses on phase-relevant sensory cues. The project page says it works through attention masking without modifying the input structure.

What is L-CALVIN?

A benchmark the abstract says the authors propose to evaluate long-horizon manipulation. How it is built is Not stated in the text read.

How is Long-VLA different from MemoryVLA?

They are different papers. The site entry describes MemoryVLA as adding a perceptual-cognitive memory bank; Long-VLA uses phase-aware input masking. The sources read do not compare them.

Is Long-VLA code available?

The GitHub repository found, long-vla/long-vla.github.io, holds the project page source. No code or checkpoint release was found, so availability and licence are Not stated.

Does Firsthand use Long-VLA, and can I request a custom dataset?

Firsthand is not affiliated and claims no use, training or compatibility. The site states one task instruction per episode and ordered action segments, with phase labels Not stated. You can request a custom dataset, confirmed first on a Spec Pilot.