Efficient VLAs

VLA: Vision-Language-Action Model

Overview

Efficient VLA research studies how robots and other embodied agents can connect visual observations and language instructions to actions under limited compute, memory, and latency budgets. Common approaches include compact backbones and action heads, visual-token pruning, quantization and cache reuse, and faster action generation through chunking, parallel decoding, or fewer diffusion steps. Other methods adapt computation to the task or reuse compressed memory so models can respond quickly while retaining context for long-horizon manipulation, navigation, and driving. Results on benchmarks such as LIBERO, RoboCasa, and driving and real-robot evaluations show that many methods can improve speed or reduce resource use, but efficiency must also be judged by task success, robustness, motion quality, and end-to-end execution cost.

Latest papers

110 papers9 papers in the last four weeks, down 50% on the four weeks before. 0.1% of all new papers.

Nov 18, 2026

Allocentric Memory, Egocentric Acting: A Dual-Frame VLA for Urban Aerial Navigation

Language-goal aerial navigation requires a UAV to reach a destination specified in natural language using observations collected during flight. A key challenge is linking changing egocentric perception with spatial evidence anchored to stable locations for long-range navigation. We propose CityNav-VLA, a dual-frame vision-language-action model for urban navigation integrating dynamic egocentric context with persistent allocentric spatial memory. The UAV incrementally accumulates visual observations on a fixed north-up map initialized with geographic and landmark priors, preserving evidence across viewpoint changes. A dynamic query conditioned on the instruction, current observations, UAV state, and motion history retrieves relevant memory evidence to predict a map-frame target. A deterministic controller converts this target into body-frame commands using the UAV's current pose, connecting global reasoning with egocentric execution. To reduce target fluctuations over long horizons, early goal commitment (EGC) fixes the execution target to the coordinate-wise median of the first KK predictions without extra trainable parameters. On Test Unseen, CityNav-VLA reduces NE by 21.2% and improves SR and SPL by 15.8% and 30.9%, respectively. It also reduces NE by 6.2% and improves SR and SPL by 5.6% and 23.2%, respectively, relative to the former SOTA after dataset refinement. Cumulative ablations show that navigation-map features yield the largest gains, while metric coordinates further improve endpoint accuracy. With a 4B-scale backbone, inference averages 0.68 s per decision with 8.6 GB peak memory on a single A100, presenting remarkable model efficiency.
Nov 17, 2026

CoBoVLA: Decoupling Language from Action with Discrete Codes for Robust and Interpretable Robot Manipulation

Vision-Language-Action (VLA) models achieve high success rates on the LIBERO benchmark but remain sensitive to paraphrasing. Existing VLAs ranging from 0.6B to 7.5B parameters experience a 22-52 point performance drop on LIBERO-Para while SmolVLA exhibits an even larger decline, from 83% to 8%. Because most VLAs jointly learn language understanding and action generation, it is difficult to determine whether a failure originates from misunderstanding an instruction or executing it incorrectly. To address these challenges, we introduce CoBoVLA, or CodeBookVLA, a 250M-parameter VLA that decouples language understanding, planning, and control through learned discrete interpretable codes. A text encoder maps the instructions to shared step codes, which a visual planner then converts into action code sequences executed by a language-independent actor. This design improves robustness to instruction paraphrasing while allowing failures to be traced to individual stages. It also enables efficient real-time control on resource-constrained hardware by encoding instructions only once and using compact discrete representations to keep each component lightweight. CoBoVLA achieves 91.4% success on LIBERO-Goal, compared with 82.8% for SmolVLA, under a matched protocol, with 17 ms per control step latency. On LIBERO-Para, it achieves 81.1% across 4,092 unseen paraphrases and 78.5% on matched episodes, compared with 36.5% for SmolVLA LoRA fine-tuned on paraphrases. Without two-step demonstrations, CoBoVLA completes 48.8% of 160 free-form two-step tasks. Finally, success reaches 91% when instructions are correctly encoded but falls to 27.2% otherwise, identifying language understanding as the main source of remaining failures. Together, these results demonstrate that decoupling language understanding and control can improve instruction robustness while making VLA failures easier to diagnose.
Nov 11, 2026

GraSLA: A Lightweight Spatial-Language-Action Model with Grasp Hypotheses as Geometric Priors

This paper focuses on improving the effectiveness and generalization of language-conditioned manipulation policies learned through imitation learning. Recently, continuous-control policies have become mainstream in robotic manipulation, but the lack of diverse demonstrations tends to restrict their generalization. To address this limitation, explicit geometric targets can provide additional guidance for action prediction. In particular, 6-DoF grasp detection generates diverse grasp hypotheses that suggest possible approach directions for the gripper. This motivates us to integrate them into continuous-control policies. To this end, we propose GraSLA, a lightweight and effective spatial-language-action model with grasp hypotheses as geometric priors. In particular, we employ spatial and language foundation models to encode scene geometry and task semantics, and integrate them with grasp hypotheses through bidirectional cross-attention for continuous action prediction. Simulation experiments demonstrate that our approach outperforms existing baselines on LIBERO, LIBERO+, and LIBERO-Pro, achieving average success rates of 99.30%, 73.7%, and 53.33%, respectively.
Nov 7, 2026

Spatial Representation Distillation and Knowledge Routing for Vision-Language-Action Models

Understanding three-dimensional physical structure has become essential for vision-language-action models (VLAs). However, existing approaches either rely on explicit geometric inputs at inference time or adopt a single visual pathway, potentially compromising the visual-semantic knowledge inherited from pre-trained models. We present SPARK-VLA, Spatial Representation distillation and Knowledge routing for VLAs, a preserve-specialize-route framework for spatially adapting VLAs without requiring external 3D inputs or experts at inference. SPARK-VLA consists of two key components: (1) spatial representation distillation, which transfers 3D understanding from a frozen 3D expert to a LoRA-adapted visual branch by aligning attention responses while preserving the original visual pathway, and (2) a lightweight knowledge router, which selectively forwards action-relevant visual tokens to the language model using routing supervision derived from action-loss gradients. Across LIBERO, LIBERO-Plus, CALVIN, and real-world manipulation, SPARK-VLA achieves strong performance with only a 0.5B language backbone and approximately 53M trainable parameters, providing an effective and parameter-efficient approach to 3D-aware VLA learning. Code is available at https://github.com/park-jungin/spark-vla.
Oct 29, 2026

OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation

Human cognition flexibly draws on a shared perceptual foundation to understand environments, attend to referenced entities, reason about spatial relationships, and act. Endowing aerial robots with an analogous functional organization remains challenging, as existing systems are often specialized for individual capabilities, dependent on offboard computation, or unable to sustain closed-loop operation on resource-constrained platforms. We present OpenFlyV2, a unified onboard vision-language-action architecture that uses a shared vision-language representation to support semantic understanding, language-grounded perception, spatial reasoning, and continuous aerial control within a single model. A compact action expert converts the shared multimodal context into short-horizon flight waypoints, while motion-aware temporal memory and action priors reduce the computational cost of closed-loop execution. We construct \dataset{}, a simulation-based dataset comprising approximately 70,000 navigation trajectories and 60,000 vision-language instruction--response pairs. OpenFlyV2 generates action chunks, each comprising 12 waypoints, entirely onboard at 10 Hz using an NVIDIA Jetson Orin NX. It transfers zero-shot from simulation to physical quadrotor flight, achieving a 48% navigation success rate across four real-world environments without real-world aerial training data. Extensive results demonstrate that broad vision-language competence and continuous closed-loop flight can coexist within a single model on a resource-constrained aerial platform. Code will be made publicly available at https://github.com/Ideal-ljl/AeroBrain.
Oct 28, 2026

Quantize by Doing: Task-Guided Mixed-Precision Quantization of Vision-Language-Action Models

Post-training quantization (PTQ) reduces the memory footprint of vision-language-action (VLA) models, but lowering bit widths while maintaining task performance remains challenging. Existing work has used sensitivity estimates from calibration data to assign bit widths to different model components. However, using task success directly to guide mixed-precision allocation is underexplored because closed-loop evaluations are costly and the effect of quantizing one component depends on the precision of other components. To this end, we propose QbD-VLA, a task-guided mixed-precision quantization method that uses task success from short closed-loop evaluations to assign weight and activation bit widths while accounting for interactions between model components. Experiments covering four VLA models and all four LIBERO suites show strong task performance at average precisions of approximately 4 bits for weights and fewer than 6 bits for activations. On SIMPLER, the gain over uniform quantization is small. On a Franka Research 3 robot arm, simulation-selected allocations show no detected success-rate loss on X-VLA but fail on π0.5π_{0.5}. These findings support using task success to guide mixed-precision VLA quantization, while highlighting the need for real-world evaluation before deployment. Code is available at github.com/jiuyixu25/QbD-VLA.
Oct 27, 2026

Vision-Language-Action or World-Action? Just Action Models

The vision-language-action (VLA) and world-action model (WAM) communities debate whether control policies should inherit pretrained knowledge from vision-language models or from visual generators. The debate, however, bundles three separable choices: backbone initialization, the objective applied to future observations, and whether the machinery for that future loss is kept at inference. A gap between a VLA and a WAM therefore cannot simply be attributed to any one of them. We show that decoupling these choices yields a common framework for both families, which we term Just Action Model (JAM): a pretrained visual backbone with an action expert, trained with an action loss and a foresight loss. The foresight loss trains current-observation features to predict frozen encodings of the recorded future through a prediction head discarded after training, so no visual generation is involved in either training or inference, and the backbone can be chosen freely. We first compare objectives within fixed policy architectures, then evaluate the foresight recipe across backbone families. The resulting JAM policies outperform their matched FastWAM and ImageWAM counterparts on LIBERO and LIBERO-Plus, and achieve similar or slightly higher success rates on RoboTwin. Compared with FastWAM, JAM improves the success rates from 97.0%/55.3%/91.2% to 98.2%/64.5%/91.4%, respectively. With cached foresight targets, JAM reduces estimated online training FLOPs per sample per update by 73% relative to FastWAM. Our results show that future supervision improves control without visual generation or additional inference machinery, and strong foresight performance is attainable with either generator-initialized or vision-language backbones. Project Page: https://y0uroy.github.io/JAM
Oct 22, 2026

Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates

When a vision-language-action (VLA) model is fine-tuned with its trainable weights in bfloat16 (bf16) and no fp32 master copy, AdamW writes each update straight into bf16, silently rounding away any update smaller than 1/512 of a weight. In 9 of 12 open VLA codebases a default or documented recipe trains this way, 4 on full weight matrices, including LeRobot's SmolVLA default and its π0π_0/π0.5π_{0.5} guides. In our LIBERO runs without the copy, a probe every 100 steps finds at least 79% of updates to bf16 weights erased. On the Hugging Face Hub, 99% of sampled full-weight SmolVLA fine-tunes store the action expert in bf16, and 84% of recent π0.5π_{0.5} fine-tunes with a training configuration set the bf16 flag. On LIBERO-Object, with rules fixed in advance, an fp32 copy raises success over LeRobot's storage from 4.3% to 49.3% for SmolVLA and from 5.1% to 80.9% for π0.5π_{0.5} at a ten-times lower learning rate, where both models are undertrained, and from 89.0% to 97.7% for π0.5π_{0.5} at its default rate. For SmolVLA at its default rate it adds 10.2 points with open-loop 50-action chunks but only 4.2 when re-planning every 10 actions. In these five comparisons and two more against SmolVLA with every trainable weight in bf16, every run with the copy beats every run without it (p < 0.05); all gaps but the 8.7- and 4.2-point ones clear our 10-point threshold. The fix is to keep the trainable weights in fp32 during training, a one-line change for SmolVLA in LeRobot.
Oct 22, 2026

SmolVLA-MoE: Sparse Mixture-of-Experts for Efficient Multi-Task Robotic Manipulation

Vision-Language-Action (VLA) models have shown promising results in multi-task robotic manipulation by leveraging visual and language understanding for action generation. However, existing lightweight VLA models employ a single dense action expert to handle all tasks, which limits their multi-task capability under constrained computational budgets. To address this issue, we propose SmolVLA-MoE, a novel framework that introduces sparse Mixture-of-Experts (MoE) into the FFN layers of the action expert in VLA models. SmolVLA-MoE replaces dense FFN layers with multiple parallel expert FFNs and a token-level router, enabling each token to adaptively select the most suitable expert for action generation. We further introduce a shared expert with a tunable weight parameter to balance the inheritance of pre-trained capabilities and task-specific expert specialization. Through sparse activation, SmolVLA-MoE activates only a single expert FFN during inference, with a total activated parameter count of approximately 500M including the frozen VLM backbone, while benefiting from approximately 800M total stored parameters. We evaluate our method on four task suites of the LIBERO benchmark. SmolVLA-MoE achieves an average success rate of 84.25%, outperforming SmolVLA (82.75%) and OpenVLA-7B (76.5%), while maintaining comparable inference efficiency. Our code is publicly available at https://github.com/dudu435/smolvla-moe.
Oct 19, 2026

CondenseVLA: Learnable History Condensation for Efficient Multi-Frame VLA

Robot manipulation often depends on short-term cues - such as brief contacts, temporary occlusions, and intermediate object states - that are ambiguous in a single frame and can only be resolved by integrating observations over time. However, injecting historical information solely by enlarging the history window is computationally costly and frequently yields limited gains due to substantial redundancy and task-agnostic, inflexible sampling of observations. We propose CondenseVLA, a multi-frame VLA policy with a learnable decoder-side history condensation module for using short-term visual context in action decoding. Given a sequence of cached per-frame features from the VLM backbone, CondenseVLA distills observation history into a fixed set of slot tokens using learnable queries and multi-head cross-attention. Simple auxiliary regularization penalizes overlap among slot attention maps and decorrelates slot embeddings, encouraging complementary attention and reducing slot redundancy. Across the evaluated tasks, CondenseVLA improves Simpler-WidowX success from 58.7% to 69.8% (+11.1 percentage points), achieves 97.2% on LIBERO, and reaches 52.8% on real-world tasks. It achieves 21.6% success on memory-dependent RMBench tasks. These results support selective condensation of bounded recent visual evidence for action decoding, while harder multi-observation tasks expose the limits of a finite history window. Project page: https://github.com/FeiyangHong/CondenseVLA
Oct 19, 2026

Think Densely, Act Sparsely: Latent Expert Cognitive Chains for Vision-Language-Action Autonomous Driving

Autonomous driving fundamentally demands dense understanding of visual, semantic, and geometric information, yet existing Vision-Language-Action (VLA) models are typically trained with sparse supervision such as language instructions or trajectory signals. This direct mapping from dense visual observations to sparse representations forces the model to compress or even discard critical scene information essential for safe driving. We argue that this limitation does not stem from insufficient model capacity, but from the lack of explicit mechanisms that encourage the formation of dense world representations in the latent space. In this spirit, we propose a novel paradigm for driving intelligence, Think Densely, Act Sparsely, and instantiate it with an efficient and interpretable framework, LECDrive (Latent Expert Cognitive Chains for Vision-Language-Action autonomous Driving). The core idea is to mimic the hierarchical perception process of humans by embedding a progressive chain of latent expert cognition within the VLA model. Concretely, we inject a compact set of task-specific latent tokens as carriers, enabling the model to internalize dense knowledge from foundational vision models, including visual representations (DINOv3), semantic structures (SAM3), and spatial geometry (DepthAnything3), while maintaining efficient inference. This interdependent cognitive chain follows a curriculum learning principle, where supervision is progressively structured from representation to semantics and geometry. Extensive experiments demonstrate that LECDrive consistently improves the performance of base VLMs across multiple autonomous driving benchmarks, covering key tasks such as object perception, state prediction, and trajectory planning. These findings suggest that constructing dense world models through structured latent expert chains is a promising direction for VLA-based autonomous driving.
Oct 16, 2026

NestedVLA: Learning to Consolidate and Generate Skills for Vision-Language-Action Models

Vision-language-action (VLA) models can learn a wide range of robotic skills, but combining separately trained skills remains challenging. The original training data may be unavailable for joint retraining, while directly merging model weights can cause interference. We propose NestedVLA, which consolidates independently fine-tuned VLA policies without access to their original training data. Instead of merging all updates into one checkpoint, a hypernetwork embedded in the VLA generates action-expert weight updates for a selected task. Each task's compressed vision-language model update is stored separately, and both updates are added to a frozen base VLA to instantiate the corresponding policy. We evaluate NestedVLA by consolidating three policies fine-tuned with reinforcement learning on MetaWorld, ManiSkill, and CALVIN. It achieves 66.7% average success, improving on the strongest model-merging baseline by 41.1 percentage points. It performs within 1 percentage point of per-task fine-tuning on ManiSkill and substantially narrows the gap on CALVIN. In this three-policy setting, total storage is 1.5x that of a single full checkpoint. Code is available at https://github.com/WanyuGroup/NeurIPS2026-NestedVLA.
Oct 15, 2026

REKI: Registering Knowledge in VLAs

We introduce REKI, a 2.4B vision-language-action model (VLA) pretrained exclusively on public data that sets a new Pareto frontier between performance and model size: on LIBERO-Plus, RoboTwin Clean2Rand and RoboCasa365, it is outperformed only by models about twice its size. REKI consists of a vision-language model (VLM) followed by a flow-matching action expert, and is pretrained on a curated mixture of public robot, egocentric human and vision-language data. Additionally, we endow it with a memory that gives the policy access to its recent past at negligible cost: a few register tokens in the action expert summarise the output of the VLM at each frame and persist in the expert's cache. This memory improves success in simulation and on real robots, and makes oscillations between poses about four times rarer. REKI uses a simple model design and training paradigm, and is optimised to be efficient at training and inference. We release the full code across data curation, training, inference and evaluation, deployment, and all pretrained and finetuned model weights: https://github.com/kyutai-labs/reki.
Oct 11, 2026

Vision-Predictor-Action: Towards Efficient Embodied AI via Neuro-Symbolic JEPA Predictors

Vision-Language-Action (VLA) models commonly generate robot actions token by token through Large Language Model (LLM) decoders, causing inference latency to grow with action sequence length and limiting closed-loop control frequency. We present Goal-Conditioned Vision-Predictor-Action (VPA), which replaces autoregressive decoders with a lightweight latent control loop. A shared vision encoder and a frozen text encoder embed observations, goal images, and instructions; a neuro-symbolic selector chooses a discrete skill; a Joint-Embedding Predictive Architecture (JEPA) predictor forecasts the next latent state; and a flow-matching solver generates an entire action chunk in a few integration steps, so that sequential network evaluations per decision step do not depend on the action horizon. We analyze inference complexity, regularization against representation collapse, and conditions under which latent sub-goal thresholding limits error accumulation across task phases. As an initial proof of concept, we evaluated VPA on the LIBERO-10 suite. The trainable model has under 20M parameters (checkpoint under 75 MB, excluding a frozen 63.4M-parameter CLIP text encoder); a decision step of a single-camera reference configuration takes 10-14 ms on an Apple M4; and trained only on LIBERO-10 demonstrations in under three hours on a single GPU, a two-camera policy succeeded in 40.0% of 200 evaluation episodes. These preliminary results from a single training run suggest the approach is workable. Code and weights are available at https://github.com/ylee-sling/Vision-Predictor-Action and https://huggingface.co/ylee-sling/vpa-libero10, and archived at https://doi.org/10.5281/zenodo.22960920.
Oct 9, 2026

Evo-World: A Lightweight Vision-Language-Action Model with Future Motion Prediction via Optical-Flow Supervision

Vision-language-action (VLA) models provide a unified framework for visual understanding, language grounding, and robot control, yet anticipating how the scene will evolve beyond the current observation can provide additional information for control. Action supervision alone does not explicitly encourage such anticipation, while future-image prediction entangles task-relevant motion with largely static appearance. To address these limitations, we introduce Evo-World, a lightweight VLA model that augments policy learning with future motion prediction through optical-flow supervision, thereby providing a task-relevant predictive signal focused on motion rather than appearance. We further design the training procedure to preserve these predictive representations during policy learning, regulating representation diversity while decoupling representation learning from action optimization. Across simulation benchmarks, Evo-World demonstrates strong performance, achieving a tier-averaged success rate of 91.1% on Meta-World MT50 , with a 13.0% relative improvement over the non-predictive baseline. On real-world AgileX robots, Evo-World outperforms the evaluated baselines across single-arm and bimanual manipulation tasks while using only 0.87B parameters and delivering lower GPU memory usage and higher inference efficiency. These results demonstrate that future motion prediction through optical-flow supervision provides an effective and lightweight approach to enhancing the predictive capabilities of VLA models.
Oct 8, 2026

Tell Robot What Not to Do: A Negation Understanding Perspective

Instruction following enables robots to perform diverse tasks specified in natural language, making it a fundamental capability for human-robot interaction. Beyond communicating desired outcomes, users also need to specify constraints on what not to do. We investigate how to enable vision-language-action models (VLAs) to follow negated instructions, where robots must accomplish task goals while respecting explicit exclusions. To this end, we propose NegaAlign, a parameter-efficient, plug-and-play framework that extends pretrained VLAs to follow negated instructions through image-language supervision alone. Specifically, we introduce Negation Transformation Layers into selected layers of the vision-language backbone to reshape intermediate instruction representations. Meanwhile, a teacher-guided alignment mechanism is designed to align instruction-relevant visual tokens, transferring action-relevant grounding from instructions that satisfy the negated constraint. The training phase uses supervision constructed from existing demonstrations and updates only the inserted layers, keeping all pretrained parameters frozen, including the action generator. We further introduce NegaBench, a simulation benchmark spanning 10 scenarios across five domains for systematically evaluating manipulation under negated constraints. Experiments across GR00T, π0π_0, and π0.5π_{0.5} demonstrate consistent improvements in negated instruction following. With 11.6M trainable parameters, NegaAlign increases the negated-instruction success rate of π0.5π_{0.5} from 2.60% to 88.45% on NegaBench and from 12.4% to 88.8% on real-world tasks, while retaining performance on affirmative instructions.
Oct 7, 2026

NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime

Language models trained with long-horizon agentic reinforcement learning can generalize knowledge through reasoning, express precise actions, and pursue goals over many steps, raising the ceiling on what an embodied agent can understand and decide. Physical interaction, however, remains the domain of action policies, which provide dense, low-latency control. We present NavGPT-3, a harness that connects the two models, with an OS-like runtime built above it: reasoning, acting, and monitoring run as threads with their own context, tools, and permissions, while the runtime schedules them and decides which thread controls the robot's motion, so that the robot can react to sudden real-world events through interruption and thread switching. Beneath it, our action policy NavGPT VLA, trained on 19.28M examples, allocates visual tokens using codec allocation, in proportion to scene change; its 8B model alone reaches 76.94 SR on R2R-CE and leads RxR-CE with 82.39 SR. With the complete harness, NavGPT-3 sets the state of the art on R2R-CE (81.51 SR) and, for the first time, brings an autonomous agent to human level: on RxR-CE it matches human followers in success (90.43 vs. 90.4 SR) and path fidelity (78.47 vs. 77.7 nDTW) at 1 min 22 s per episode, versus roughly 3 min for a human. We comprehensively ablate the harness design and the interaction between the two models, showing how tools and the action policy shape the path from language-model reasoning to physical control: when NavGPT VLA executes the route, the reasoning loop shortens and the system's minimum reaction time falls from 3-19 s per language-model decision to 0.5-1 s per action-policy step (1-2 Hz). These results show that designing this embodied interface is central to connecting frontier language-model intelligence with low-level physical control. We will release all models, code, and evaluation records.
Oct 6, 2026

VLA-ACL: Action-Consistent Visual Token Pruning for Efficient Vision-Language-Action Models

Vision-Language-Action (VLA) models achieve strong robotic manipulation performance but incur high computational costs from processing long token sequences at every control step, limiting real-time deployment. Visual token pruning offers a direct solution, as visual patches dominate the input sequence and contain considerable redundancy. Existing approaches, however, either rely on indirect training-free heuristics, such as attention scores and motion thresholds, or require costly fine-tuning of the base VLA model. We introduce VLA-ACL (Action Consistency Learning), which learns a lightweight visual token pruning policy through action-level supervision while keeping the base VLA model entirely frozen. The training objective encourages actions produced from pruned visual contexts to remain consistent with the full-context teacher, with ground-truth actions as auxiliary supervision. This directly ties token selection to its effect on the downstream control output. Experiments on LIBERO and real-world manipulation tasks show that VLA-ACL prunes up to 87.5% of visual tokens while retaining competitive performance, reduces computation by up to 75%, and achieves a 1.5x inference speedup. These results establish a stronger performance-efficiency trade-off than existing frozen-VLA pruning methods and demonstrate the value of action-level supervision for visual token selection. Code is available at https://github.com/du-owen/VLA-ACL.
Oct 2, 2026

GeoScaffold: Learning Compact Geometric Latents via Reconstruction for Efficient Vision-Language Navigation

Recent vision-and-language navigation (VLN) systems increasingly adopt streaming Video-LLM policies that map egocentric RGB observations and instructions directly to low-level actions. Yet these policies inherit weak 3D geometric priors from 2D pretraining. Existing geometry-aware extensions charge a persistent inference-time price: depth sensors, 3D encoders, or per-step perception tool calls. We propose GeoScaffold, a geometric supervision framework that pays this price once, at training time, by internalizing geometry into the policy itself. It first learns a compact depth tokenizer on depth maps from the training trajectories and freezes it. It then fine-tunes the policy with a handful of learnable geometry query tokens, training their hidden states to reconstruct navigation-critical geometry such as depth, connectivity, and traversability. This supervision turns the query states into compact geometric latents for action decoding, and through the shared weights also internalizes geometry into the backbone's own representations. Like a scaffold, the tokenizer, target generators, and reconstruction heads are discarded after training, leaving the backbone and action interface unchanged. Extensive experiments show that GeoScaffold consistently outperforms leading vision-only navigators on continuous VLN benchmarks, offering a practical paradigm for lightweight edge deployment of spatially aware embodied navigation models.
Oct 1, 2026

eRLT: Efficient VLA Reinforcement Learning via Action-Relevant Token Routing

Vision-Language-Action (VLA) models provide strong behavioral priors for robotic manipulation, yet efficiently adapting them to downstream tasks remains challenging. Recent work addresses this challenge by adapting frozen VLAs through online reinforcement learning (RL), whose sample efficiency depends on the quality of the state representation used by the actor and critic. Existing methods construct such representations either with VLA-independent visual encoders or through fixed compression of internal VLA representations. Neither design explicitly extracts the task-specific action-relevant VLA features most useful for downstream action refinement and action-value estimation, therefore limiting sample efficiency. To address this limitation, we introduce eRLT, which constructs an effective state representation by routing task-specific action-relevant information across both tokens and layers of the frozen VLA. Specifically, learned routing tokens dynamically aggregate visual-language features at multiple depths, while a lightweight layer router combines these summaries into a fixed-dimensional RL token. The routing module is initialized using expert demonstrations to capture features predictive of expert actions and then refined using critic feedback from online interactions for action-value estimation. Across seven LIBERO and RoboTwin tasks, eRLT improves mean normalized learning-curve AUC by up to 23.7% over representative baselines. Real-robot experiments on USB connector insertion and motherboard ribbon-cable insertion further show AUC improvements of 108.9% and 46.7%, respectively, over the strongest baseline. Project page: https://red0orange.github.io/eRLT_site/.
Sep 30, 2026

Spike-driven Vision-Language-Action Model

Vision-language-action (VLA) models bridge multimodal understanding and robotic control, advancing the dominant paradigm for embodied intelligence. However, most existing models rely on large Transformers, whose latency and energy costs hinder deployment on resource-constrained platforms. Through sparse event-driven computation, spiking neural networks offer a promising paradigm for high-performance and energy-efficient computing. Here, we propose the first Spike-driven VLA framework enabling end-to-end direct training for robotic manipulation, which mainly comprises three core components. First, we develop spiking visual and instruction encoders for multimodal perception, encoding visual observations and language instructions into sparse, reliable spike representations for subsequent cross-modal fusion. Then, we introduce Multi-Winner Spike Fusion for instruction-guided scene understanding, using bidirectional top-kk winner-take-all spike routing to suppress background interference and yield fused memory. Finally, we propose a Spike Action Chunking Transformer that incorporates spiking cross-attention over the fused memory and the current robot state, enabling efficient end-to-end generation of continuous action chunks for robotic control. Extensive experiments on LIBERO and Meta-World demonstrate that Spike-driven VLA achieves competitive performance with fewer parameters and lower estimated inference energy than conventional VLA models. This work establishes a foundational framework for neuromorphic VLA modeling, paving the way for future advances in resource-efficient embodied intelligence.
Sep 28, 2026

ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning

Vision-Language-Action (VLA) models have shown great promise for robotic manipulation by mapping multi-modal semantics to physical actions. However, this mapping inherently struggles to align these coarse-grained semantics with fine-grained temporal execution. It leaves VLA models with limited generalization and insufficient robustness in cluttered environments. To overcome this issue, we propose ActionUNet, an efficient multi-scale fine-tuning framework that enhances pre-trained VLA models with minimal computational cost. ActionUNet first constructs a lightweight temporal U-Net within the temporal-aligned action feature space to fuse hierarchical structural priors, effectively bridging the scale gap between semantics and temporal executions. Recognizing that multi-scale modeling can disrupt microscopic temporal continuity and cause mechanical oscillations, ActionUNet then employs a conditional SIREN as a continuous action decoder. Equipped with explicit second-order smoothness constraints, this decoder guarantees temporal continuity and reduces high-frequency motion jitter. By smoothing temporal discontinuities from multi-scale fusion, this continuous formulation reduces mechanical execution failures while preserving the base VLA model's generalization and manipulation robustness. Extensive experiments on RoboTwin 2.0 and LIBERO-Plus benchmarks, together with real-world hard evaluations, demonstrate that ActionUNet significantly improves π0.5 success rates by absolute 9.8%, 6.1%, and 11.4%, respectively, while also generalizing to the regression-based OpenVLA-OFT backbone, highlighting its effectiveness and efficiency as a fine-tuning strategy. Code and implementation details are available at https://github.com/Di-Zhu123/ActionUNet.
Sep 28, 2026

Adjoint Guidance Flow: Amortized Critic Guidance for VLA Policies

Flow-based Vision-Language-Action (VLA) policies are typically trained by behavior cloning and thus do not explicitly optimize long-term task return. Critic guidance steers generation toward higher-value actions, but existing methods differentiate the critic through a one-step surrogate of the sampler and back-propagate a critic ensemble at every flow step. In contrast, here we propose Adjoint Guidance Flow (AGF), which amortizes trajectory-aware critic guidance into a lightweight guidance network while preserving the pretrained VLA policy. Specifically, we formulate critic-guided flow generation as a deterministic optimal control problem, whose optimal guidance is a costate that carries the terminal critic gradient back through the remaining flow, and regress the guidance network onto this costate while keeping both the VLA and critic frozen. This design provides favorable memory and throughput scaling during training, and inference needs one guidance-network forward pass per step, without the critic ensemble, back-propagation, or adjoint computation. Across LIBERO, RoboCasa, and LIBERO-Pro, AGF consistently improves pretrained VLAs, remains competitive with critic-guidance and policy-fine-tuning baselines, and is the most robust method when a single guidance strength is deployed across tasks. Compared with QGF, AGF runs 3.6×3.6\times faster per guidance step with 7.0×7.0\times fewer parameters, with comparable and even better performance, showing that critic guidance can be trajectory-aware and lightweight.
Sep 28, 2026

FailPatch: Failure Residual Patching for Vision-Language-Action Models

Vision-Language-Action (VLA) policies are typically adapted using successful demonstrations, which provide direct action supervision but rarely cover failure-prone states. Deployment failures expose these states, yet lack the corrective actions needed for conventional supervised learning. We propose FailPatch, a failure-driven residual patching framework that decouples action supervision from execution-reliability supervision. Successful demonstrations ground how the policy should act, while deployment trajectories indicate when its behavior becomes unreliable. We further observe that action hidden representations exhibit clear linear separability between reliable and failure-associated states while directly conditioning action generation. Building on these insights, FailPatch introduces a Null-gated Residual Expert Bank into the action hidden space of a frozen VLA policy. A unified Preserve--Redirect--Trust objective retains the original policy in reliable states, selects residual experts in failure-associated states and redirects representations from failure regions toward success-associated regions under bounded intervention. With only 0.52% trainable parameters, FailPatch improves success rates by 11.0 percentage points on four long-horizon RoboTwin tasks under clean evaluation, 9.5 percentage points under clean-to-random generalization, and 16.7 percentage points over the baseline across three real-world tasks. Project and code: https://github.com/yupeng-2003/FailPatch.
Sep 27, 2026

SLIP-VLA: Single-Step Latent Imagination for Policy Learning in Vision-Language-Action Models

Vision-Language-Action models are increasingly effective for robotic manipulation, yet most predict actions directly from current observations without explicitly modeling future scene evolution. Recent methods introduce future prediction to improve action generation, but dense future modeling often requires expensive iterative denoising, while one-step alternatives can underperform their multi-step counterparts. To reconcile efficient future modeling with strong action performance, we present SLIP-VLA, a policy learning framework that equips VLA models with a Single-Step Latent Imagination for future-aware action prediction. SLIP-VLA obtains temporally dense future latent representations with a single denoising update, and we improve the perceptual sufficiency of these representations by aligning intermediate latents with future geometric and semantic features. We further improve their control sufficiency through action-conditioned latent world modeling and inverse dynamics modeling, explicitly coupling latent transitions with robot actions. SLIP-VLA achieves state-of-the-art performance across diverse simulation benchmarks and real-world manipulation tasks, while its single-step latent imagination takes only 12 ms.
Sep 25, 2026

Towards VLA-Dreamer: Refining VLA Behavior Using World Models

Vision-Language-Action models (VLAs), while showing strong potential for robot control, require massive amounts of high-quality imitation learning data. Moreover, the absence of an explicit world model casts further doubt on their control capabilities. In this concept paper, we propose a novel architecture that addresses sample efficiency in VLAs by training a predictive world model on the embedding space of the VLA's vision encoder. We hypothesize that these embeddings are action-relevant and usable for future prediction. To this end, we propose using the suggested architecture to investigate how well these embeddings predict the future based on actions, as the inability to do so would mark a key limitation of VLA architectures: the lack of a non-lossy implicit world model to simulate real-world dynamics. The proposed architecture differs from the standard world model dynamics as the loss comes from the embedding space rather than the pixel space, similar to joint embedding predictive architectures. Furthermore, the trained world model can be utilized for short-term planning tasks by sampling VLA actions given goal images. We intend to examine the richness of vision embeddings in VLAs and reduce their high data requirements through a world model that can also generate plans during inference.
Sep 25, 2026

FRAM: Trajectory-Guided Visual Feature Selection for Compact Language-Conditioned Robot Manipulation

Vision-Language-Action models achieve strong performance in robot manipulation, but often require large numbers of parameters. In this work, we propose the Future Representation Action Model (FRAM), a small policy that explicitly links the future end-effector trajectory to the current visual input. FRAM uses the image coordinates of the predicted trajectory as spatial pointers and reads local visual features related to the motion from the current image. This organizes the information for action generation into the reference position (Where), the visual state (What), and the future motion (Future). Trajectory labels are generated automatically from demonstrations and camera geometry, so no manual annotation is needed. With 138.7M parameters, including a frozen language encoder, FRAM reaches an average success rate of 92.2% over the four standard LIBERO suites, close to the 94.2% of π0π_0 with 3.3B parameters. Without extra training, it also reaches an average of 67.3% on LIBERO-Plus. Ablations confirm that both the future trajectory and the local visual features improve performance and robustness. On a real dual-arm UR5e, FRAM stacks cups using only wrist cameras, including choosing and switching between the left and right arms. These results show that selecting visual information based on future motion is an effective way to obtain both high performance and robustness in a small robot policy.
Sep 25, 2026

VLALight: Lightweight Vision-Language-Action Models for Emergency-Aware Traffic Signal Control

Traffic signal control (TSC) is essential for mitigating urban congestion. Recent advances in vision-language models (VLMs) enable richer interpretation of intersection scenes, opening new opportunities for visual-context-aware TSC. However, the loose coupling and repeated information conversion between modules can lead to the loss of fine-grained visual details, while sequential inference introduces substantial latency. To address these limitations, we propose VLALight, a lightweight end-to-end vision-language-action framework that directly maps intersection observations and signal-phase information to discrete signal actions. To handle the multi-view nature of TSC, VLALight combines multiple directional camera views into a unified visual input and uses textual instructions to establish their correspondence with traffic movements and signal phases. This design enables direct action prediction with a compact 0.5 B-parameter model, without intermediate image-to-text descriptions or handcrafted traffic-state representations. Experiments show that VLALight delivers the best emergency-vehicle service of all compared methods, reducing pooled emergency waiting time by 21.1% over the cascaded VLMLight while running in real time on local hardware and generalizing to unseen intersection topologies and traffic-flow patterns.
Sep 23, 2026

TANDEM: Task and Motion Planning with As-Needed Demonstrations for Efficient Vision-Language-Action Model Fine-tuning

Human teleoperators spend substantial time demonstrating behaviors that robots can already perform autonomously, limiting the scalability of data collection for robot foundation models. Task and motion planning (TAMP) can automate many of these behaviors, but a fixed planning domain may not support every stage of a long-horizon manipulation task. We present TANDEM (Tamp with As-Needed Demonstrations for Efficient Model fine-tuning), a system that combines TAMP with selective human teleoperation to collect demonstrations for tasks beyond the planner's capabilities. Our key idea is to represent human assistance as an on-demand planning capability. Given a language instruction and visual observation, TANDEM uses pretrained vision-language models to extend the planning domain with missing predicates and human-executed magic operators. This allows the planner to interleave autonomous and human-executed stages without task-specific intervention points. After each human stage, TANDEM re-perceives the scene and checks whether the intended effects hold before resuming autonomous planning. To support fine-tuning vision-language-action (VLA) models, TANDEM also uses example pretraining trajectories to align planner-generated motions with the target model's pretraining distribution. We evaluate TANDEM on five long-horizon manipulation tasks beyond the TAMP domain's capabilities. On a representative long-horizon task, TANDEM collects 2.9x as many demonstrations as full-task teleoperation at the same human intervention time. Fine-tuning a pretrained π_{0.5}-DROID model on 20 TANDEM demonstrations per task increases average task success from 0% to 60% across the five tasks.
Sep 19, 2026

SmoLSTM: A Compact Vision-Language-Action Model with Recurrent Memory that Persists

Vision-language-action models often predict actions from only the current observation, which can leave tasks involving object occlusion or visually identical objects ambiguous without episode history. The usual countermeasure, widening the observation window, turns the horizon into a hyperparameter and lets per-step cost grow with it. We instead capture the episode in a recurrent state. SmoLSTM couples a frozen 256M-parameter SmolVLM backbone to a matrix-memory LSTM control layer in which observation tokens and action queries are unified in a single causal stream that is never reset throughout the entire episode. Recurrent-state storage is therefore O(1) in episode length. A flow-matching action head predicts chunks of 10 end-effector pose deltas and gripper commands at each control step. Our single policy, trained jointly on 7,461 demonstrations across 140 tasks and evaluated on held-out initial states, performs best, reaching 85.1% subgoal coverage and 77.5% full-task success on LIBERO-Mem with 0.04B trainable parameters, surpassing both the benchmark's own object-centric baseline and recent memory-based approaches. Resetting the recurrent state at every control step reduces full-task success to 7.0%, showing that the trained policy relies on context carried between decisions. The same model achieves 79.6% average success on standard LIBERO.