Long-VLA: End-to-End Long-Horizon VLA

• Long-VLA is a vision-language-action model that segments tasks into moving and interaction phases through phase-aware input masking.
• The approach employs unified training with ResNet-18 feature extraction, a GPT-2 style transformer, and conditional diffusion for precise action prediction.
• Long-VLA significantly outperforms previous methods on the L-CALVIN benchmark in both simulated and real-world multi-step robotic tasks.

Long-VLA is a Vision-Language-Action (VLA) model for robot manipulation that is presented as the first end-to-end VLA model specifically designed for long-horizon robotic tasks. It targets the degradation that conventional VLA systems exhibit when they move from short-horizon execution to multi-step manipulation with strong subtask dependencies, and it does so through a phase-aware input masking strategy that segments each subtask into moving and interaction phases. The method is paired with L-CALVIN, a benchmark that extends CALVIN from 5 to 10 subtasks per sequence, and is reported to significantly outperform prior state-of-the-art methods in both simulated and real-world settings (Fan et al., 27 Aug 2025).

1. Problem setting and long-horizon scope

Long-VLA is motivated by a specific failure mode of contemporary VLA systems: strong performance on short-horizon tasks does not transfer cleanly to long-horizon, multi-step robotic manipulation. In the Long-VLA formulation, the limiting factors are skill chaining challenges and subtask dependencies. The paper states that simple stage-wise decompositions do not sufficiently capture dependencies and state transitions between subtasks, and that errors or state mismatches propagate between subtasks, leading to compounding failures (Fan et al., 27 Aug 2025).

This problem formulation is consistent with a broader line of long-horizon VLA research. LiLo-VLA emphasizes combinatorial complexity, environmental sensitivity, and cascading failures in long-horizon manipulation (Yang et al., 25 Feb 2026). LoHo-Manip describes long-horizon execution as progress-dependent and brittle to compounding execution errors (Liu et al., 23 Apr 2026). S$^2$-VLA attributes long-horizon degradation to static feature fusion mechanisms that cannot adapt to different phases of task execution (Xie et al., 26 Jun 2026). Within that landscape, Long-VLA occupies a distinct position: it preserves end-to-end training while making phase structure explicit inside the policy itself (Fan et al., 27 Aug 2025).

A central design choice is that long-horizon manipulation is not addressed by splitting planning and control into separate models. Instead, the method uses decomposition into movement and interaction tuples for each subgoal while retaining unified optimization. The paper explicitly contrasts this with prior modular or two-model splits that block gradient flow and waste joint data (Fan et al., 27 Aug 2025). A plausible implication is that Long-VLA treats subtask phase alignment as the main bottleneck in long-horizon VLA execution.

2. Phase-aware decomposition and masking

The defining mechanism in Long-VLA is the phase-aware input masking strategy. Each subtask in a long-horizon trajectory is segmented into two phases:

• the Moving Phase, in which the end-effector is navigated to a target vicinity and is best perceived via the third-person static camera;
• the Interaction Phase, in which precise object manipulation is performed and is best perceived via the first-person gripper camera (Fan et al., 27 Aug 2025).

The masking mechanism dynamically enables or disables access to different sensory inputs. During moving, the model attends to the static or third-person view and not the gripper view; during interaction, it focuses on the gripper view, with the stated goal of reducing visual distribution shifts and ambiguity (Fan et al., 27 Aug 2025).

At the token level, each modality feature is assigned a binary mask $m \in \{0,1\}$. The attention mask matrix is defined as

$M_{ij} = m_i \cdot m_j,$

and the masked attention weights are

$$A_{ij} = \frac{\exp(P_{ij}) M_{ij}}{\sum_{k=1}^{N} \exp(P_{ik}) M_{ik}},$$

where $P$ is the query-key similarity matrix (Fan et al., 27 Aug 2025).

The decomposition is not merely a preprocessing heuristic. A dedicated phase identifier $S_p \in \{-1,1\}$ is appended to the action representation, explicitly signaling phase (Fan et al., 27 Aug 2025). The paper attributes three functions to this design: state alignment between subtask transitions, representation continuity across tasks, and a reduction in error propagation and dynamic coupling (Fan et al., 27 Aug 2025). In the article’s own framing, this is the mechanism by which skill chaining is made compatible with end-to-end VLA training.

The paper also reports that learnable masking experiments confirm that the model learns to preferentially activate third-person views in moving and gripper views in interaction, in accord with the manual design (Fan et al., 27 Aug 2025). This suggests that the moving/interaction split is not only architecturally imposed but also reflected in learned attention patterns.

3. Architecture and training formulation

The Long-VLA policy is written as

$T_e(a_t \mid s_t, d_t, g),$

where the action $a_t$ is predicted from the observation $s_t$, the detection input $d_t$, and a latent goal $m \in \{0,1\}$0 (Fan et al., 27 Aug 2025).

The observation $m \in \{0,1\}$1 consists of third-person static and first-person gripper camera images. The detection input $m \in \{0,1\}$2 is a fine-grained, language-conditioned object localization signal produced by Grounding DINO. The latent goal $m \in \{0,1\}$3 can be either the next observation $m \in \{0,1\}$4 or a language description, encoded by CLIP through an image or text encoder (Fan et al., 27 Aug 2025).

The feature stack is explicitly specified. Visual features from both cameras are encoded with ResNet-18. Detection features use positional encoding of bounding boxes and are fused with image features using a FiLM layer. A GPT-2-style multimodal transformer encoder receives concatenated features

$m \in \{0,1\}$5

where $m \in \{0,1\}$6 is the detection-enhanced static view (Fan et al., 27 Aug 2025).

The action decoder is a conditional diffusion model that predicts action through iterative denoising, with DDIM used for sampling. The final action token is

$m \in \{0,1\}$7

covering end-effector coordinates, orientation, gripper state, and the phase identifier (Fan et al., 27 Aug 2025).

Training uses a score matching loss for diffusion-based action prediction together with an InfoNCE goal alignment loss for consistent visual/language goal representation. The total loss is

$m \in \{0,1\}$8

The paper presents this formulation as preserving scalability and data efficiency because the masking module does not alter the model’s interface and can be implemented as a lightweight, architecture-agnostic module in any existing VLA without structural changes (Fan et al., 27 Aug 2025).

That architecture-agnostic claim is important for the method’s positioning. The paper reports successful application atop multiple VLA backbones, specifically MDT and HULC, with strong gains in each case (Fan et al., 27 Aug 2025). In this sense, Long-VLA is both a named model and a masking-and-decomposition recipe for long-horizon VLA training.

4. L-CALVIN benchmark and evaluation protocol

Long-VLA is introduced together with L-CALVIN, a benchmark designed to systematically evaluate long-horizon manipulation. L-CALVIN extends the standard CALVIN benchmark from 5 to 10 subtasks per sequence, with sequences constructed by systematic, state-consistent task selection rather than inefficient random sampling, and each step phase-decomposed (Fan et al., 27 Aug 2025).

The simulated evaluation uses the extended CALVIN environment in two settings: D$m \in \{0,1\}$9D and ABCD$M_{ij} = m_i \cdot m_j,$0D (Fan et al., 27 Aug 2025). The real-world evaluation uses two tasks:

• Sorting: placing blocks in a precise order, up to 8 steps;
• Cleaning: a multi-object, multi-action kitchen scenario (Fan et al., 27 Aug 2025).

The benchmark is designed to stress exactly the issues Long-VLA claims to solve. By increasing the sequence length and preserving state consistency, L-CALVIN makes subtask transition quality a first-class evaluation target rather than a secondary effect. The paper also reports an explicit measurement of skill chaining performance by comparing success in continuous multi-step settings against independent single-task settings, and states that previous VLA and compositional policies deteriorate rapidly when chaining, whereas Long-VLA remains steady (Fan et al., 27 Aug 2025).

This benchmark design places Long-VLA within a rapidly expanding evaluation ecosystem for long-horizon VLA research. Related work introduces other long-horizon testbeds, including LIBERO-Long++ and Ultra-Long in LiLo-VLA (Yang et al., 25 Feb 2026), VLABench and real-robot OOD tests in LoHo-Manip (Liu et al., 23 Apr 2026), and LIBERO and SimplerEnv in S$M_{ij} = m_i \cdot m_j,$1-VLA (Xie et al., 26 Jun 2026). L-CALVIN is distinctive in that it is directly built by extending CALVIN’s horizon from 5 to 10 subtasks (Fan et al., 27 Aug 2025).

5. Empirical performance and ablations

The principal simulation metric reported for Long-VLA is the average number of tasks completed in sequence. On D$M_{ij} = m_i \cdot m_j,$2D, Long-VLA achieves 4.75, compared with 2.96 for GR-1. On ABCD$M_{ij} = m_i \cdot m_j,$3D, Long-VLA achieves 8.24, compared with 5.68 for GR-1 (Fan et al., 27 Aug 2025). The paper states that the improvement grows with sequence length and that Long-VLA succeeds where previous methods degrade sharply (Fan et al., 27 Aug 2025).

In the real-world sorting task, Long-VLA achieves approximately 25% success for full-length sequences out of 8 tasks, while the base policy and other methods drop to zero after 7 tasks (Fan et al., 27 Aug 2025). In the real-world cleaning task, Long-VLA is reported to achieve approximately 226% gain over the base policy in the hardest and longest setting, and to remain robust to unseen lighting and visual distractions (Fan et al., 27 Aug 2025).

Setting	Long-VLA	Comparison
D$M_{ij} = m_i \cdot m_j,$4D avg. sequence length	4.75	GR-1: 2.96
ABCD$M_{ij} = m_i \cdot m_j,$5D avg. sequence length	8.24	GR-1: 5.68
Real-world Sorting	~25% full-length success	base policy and others drop to zero after 7 tasks
Real-world Cleaning	~226% gain over base policy in hardest setting	robust to unseen lighting and visual distractions

The paper further states that Long-VLA outperforms SOTA baselines such as To, GR-1, RoboVLMs, VLAS, OpenVLA, and UP-VLA across all benchmarks (Fan et al., 27 Aug 2025). Its ablations indicate that removing decomposition, masking, or unified training reduces performance, and that the combined system achieves the best robustness and consistency (Fan et al., 27 Aug 2025). Data efficiency is also emphasized: unified training surpasses decoupled or multi-model approaches, especially when data is limited (Fan et al., 27 Aug 2025).

These results are framed not simply as a numerical gain but as evidence for a particular thesis: phase-aware sensory routing improves subtask compatibility without sacrificing the scaling properties of unified VLA training.

6. Relation to subsequent long-horizon VLA frameworks

Long-VLA belongs to a broader research shift from short-horizon action imitation to long-horizon embodied execution. Subsequent and contemporary systems address the same regime with different organizing principles.

LiLo-VLA separates transport from interaction through a Reaching Module and an object-centric Interaction Module, and reports a 69% average success rate in simulation together with 85% average success in real-world evaluations across 8 long-horizon tasks (Yang et al., 25 Feb 2026). Critic in the Loop introduces a Tri-System architecture with a VLM brain, a VLA cerebellum, and a lightweight visual Critic, using event-driven scheduling for replanning under anomaly, success, or stagnation (Yi et al., 5 Mar 2026). LoHo-Manip decouples a task-management VLM from a trace-conditioned VLA executor, using a receding-horizon manager that predicts a done-plus-remaining language memory and a visual trace (Liu et al., 23 Apr 2026). Anticipation-VLA organizes execution around adaptive and recursive subgoal generation with a dynamic stack of subgoals (Zhang et al., 3 May 2026).

Other lines of work focus less on explicit decomposition and more on temporal or adaptive representation. HiF-VLA uses motion vectors as a compact temporal representation for hindsight priors and foresight reasoning, achieving 96.4% success on LIBERO-Long in its multi-view configuration and 4.35 average consecutive successful steps on CALVIN ABC-D (Lin et al., 10 Dec 2025). S$M_{ij} = m_i \cdot m_j,$6-VLA introduces belief-state-guided adaptive fusion and reports 96.4% on the LIBERO Long subset with a 2B-parameter model (Xie et al., 26 Jun 2026).

Against that background, Long-VLA is notable for keeping the VLA formulation end-to-end and architecture-agnostic while making phase structure explicit. It does not rely on a separate manager, symbolic planner, or critic. Nor does it primarily reinterpret temporal context as motion memory or a belief state. Instead, it asserts that moving and interaction phases constitute the key axis along which sensory relevance changes during long-horizon manipulation (Fan et al., 27 Aug 2025). This suggests a particular view of long-horizon control: failures in multi-step execution can be mitigated by aligning attention with subtask phase, rather than only by adding high-level planning or explicit memory.

Long-VLA is therefore best understood as an early end-to-end long-horizon VLA framework centered on phase-aware decomposition, unified training, and benchmark construction. Its importance in the literature lies both in the specific masking mechanism and in the argument that long-horizon competence can be improved without abandoning the scalability and data efficiency associated with unified VLA training (Fan et al., 27 Aug 2025).