VLA Training
VLA: Vision-Language-Action Model
Latest papers
60 papers18 papers in the last four weeks, up 200% on the four weeks before. 0.1% of all new papers.
Vision-language-action (VLA) models adapt web-pretrained vision-language models to output robot actions, so that web knowledge carries over to physical control. In under 4 years the field has moved from RT-2, the first VLA, to open models that fine-tune on a single consumer GPU and to proprietary generalists that clean unseen homes. This survey reviews 88 works from December 2022 to October 2026, mainly on manipulation and navigation, plus humanoid control, world-action models and evaluation. We follow a VLA through its life cycle, from model design to training and deployment, along seven axes. Model design covers (a) inputs and memory, what the model observes; (b) architecture, how its pipeline is built; (c) action output, how it produces motor commands; and (d) intermediate predictions, what else it predicts, from reasoning steps to future observations. Training covers (e) pretraining data, the mix of robot, web, human and simulated data, and (f) post-training, adaptation to the target robot. Deployment covers (g) real-time execution, keeping pace with the robot. We analyse 22 models on these axes and report their training data, model size, inference speed and hardware. Three trends emerge: (i) training data has broadened, from about 10,000 hours of teleoperated robot data () to 23,000 community-collected episodes (SmolVLA) and 20,000 hours of egocentric human video (GR00T N1.7); (ii) model sizes have split: most recent models have 3-7 billion parameters, while a new group under one billion runs on consumer GPUs; and (iii) generalist VLAs are emerging that handle both manipulation and navigation, such as UniVLA and Qwen-VLA. We also compare VLAs with world-action models, which predict future observations as well as actions. We review evidence that some VLAs with near-perfect benchmark scores fail when objects or the camera move slightly, and open up further research questions.
Agentic-Teacher: An Agentic Data Engine for Failure-Driven VLA Post-Training
Improving a pretrained vision-language-action (VLA) model requires demonstrations that address its weaknesses and provide effective supervision for learning. Coding agents can autonomously generate demonstrations through environment interaction, yet successful execution alone does not ensure effective supervision for VLA learning. We introduce Agentic-Teacher, an autonomous data engine for failure-driven VLA post-training. Given a VLA checkpoint, a task specification, and configured environment tools, the engine analyzes failed rollouts, selects intervention states, and iteratively develops closed-loop code policies through execution feedback. Validated policies become reusable demonstration generators that collect successful continuations from states where the VLA struggles. To make these demonstrations suitable for learning, Agentic-Teacher introduces temporally correlated action noise to broaden state-action coverage, removes idle frames while preserving gripper transitions, and filters trajectories using the frozen base model's flow-matching loss. The resulting demonstrations are combined with successful VLA rollouts for post-training, yielding an updated policy that operates without coding-agent assistance. Agentic-Teacher thus connects the weaknesses of a target VLA to both the generation and curation of its training data. In simulation, it improves success rates from 79.4% to 87.8% on LIBERO-10, from 80.4% to 90.6% on MetaWorld-10, and from 0% to 30.0% across three RoboCasa tasks.
SWAM: Scaling World Action Models with Cross-Modal Alignment
World action models jointly predict video and actions conditioned on language. Existing studies focus primarily on architectures and training mechanisms, typically treating language-video alignment acquired through video pretraining as an established foundation. However, introducing action learning may impair this alignment. Principled guidance on data and supervision design for maintaining language-video alignment during action learning remains limited. To investigate this, we curate a pool of approximately 47,000 hours of general videos, egocentric videos, UMI demonstrations, and robot demonstrations. Building on this, we systematically study how language-video supervision, language-video-action supervision, and the joint use of these two forms affect WAM performance. In particular, we examine how language annotation quality affects pretrained language-video capabilities, and how action supervision from UMI and robot demonstrations extends these capabilities to language-video-action modeling. To better scale the model, we further conduct detailed studies on joint training recipe and model architecture. Under a fixed training budget, mixing egocentric videos and UMI demonstrations improves control generalization. Model architecture affects how effectively these experiences are used, while scaling pretraining under the selected configuration brings further gains. The resulting recipe, post-trained on RoboTwin Clean demonstrations only, achieves success rates of 88.10% and 61.56% under Easy (Clean-to-Clean) and Hard (Clean-to-Random) conditions, respectively. Our model achieves 40.00% SR across three real-world tasks, compared with 23.33% for the strongest baseline, indicating better generalization under unseen conditions. We release the curated data, training code, and model weights.
From Game Generation to Game Playing: Fully Automated Ability Training for Vision-Language-Action Models
Training vision-language-action (VLA) models to follow gameplay instructions across diverse games is limited by the cost of collecting demonstrations. We introduce a framework that automatically generates game environments and expert demonstrations. An LLM-powered agent uses game-building tools to construct executable environments and tasks targeting specified skills. Programmatic experts then use privileged simulator state to generate aligned instructions, rendered observations, and keyboard and mouse actions. We train our game policy, Script2Play, by post-training a 4B vision-language model with behavior cloning on 150.3 hours of synthetic Minecraft and Roblox gameplay, without human gameplay demonstrations or reinforcement learning. Evaluation on multiple public game benchmarks demonstrates that Script2Play follows gameplay instructions effectively and generalizes across game engines, highlighting the effectiveness of targeted synthetic supervision.
SURE:Spatiotemporal Uncertainty Representation for Efficient Vision-Language-Action Learning
Vision-language-action (VLA) policies have become a promising paradigm for end-to-end robot control, but their effectiveness still depends heavily on large volumes of demonstration data. In manipulation tasks, this dependence is compounded by structured spatiotemporal uncertainty in the action space: multiple nearby actions or trajectories may successfully solve the same task, while semantically equivalent demonstrations often unfold at different speeds. Standard pointwise supervision treats each demonstrated action as an exact target, imposing an unnecessary learning burden under such uncertainty. We present SURE, a data-efficient VLA training framework that explicitly models both spatial and temporal uncertainty. SURE augments a VLA backbone with a lightweight transposed-convolutional probabilistic action decoder that captures spatial uncertainty through local correlations among neighboring action bins. It further introduces Soft-DTW alignment supervision to accommodate local timing variations between predicted and demonstrated action chunks. Together, these two components reduce sensitivity to demonstrator-specific poses and execution timing, enabling more data-efficient action learning. We evaluate SURE on LIBERO, RoboTwin 2.0, and real-world manipulation tasks. Even with only 10% of the demonstration data or 20% of the training steps, SURE achieves manipulation success rates that match or surpass baseline VLAs.
SLA: Scaling Spatial-Language-Action Model via Continual Pre-training
Reliable robotic manipulation requires both semantic understanding and precise control in three-dimensional space. However, vision-language-action (VLA) foundations largely inherit visual representations from 2D vision-language pre-training, while spatial supervision is often introduced only during task-specific adaptation. Developing spatial understanding as a transferable foundation-model capability therefore calls for learning from broader robot experience. We present SLA, a Spatial-Language-Action Model that scales implicit spatial alignment to continual pre-training and maintains it during downstream post-training. Initialized from a pretrained pi05 policy, SLA jointly optimizes action prediction and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model. We curate and unify heterogeneous robot datasets to support spatial learning across diverse embodiments, tasks, and scenes. To accommodate this diversity, we introduce feature-dependent gates that adaptively aggregate visual embeddings from multiple VLA layers for spatial supervision. The teacher further exploits partially available depth and camera calibration alongside RGB observations, enabling effective use of heterogeneous geometric annotations. Probabilistic state dropout encourages reliance on visual cues and mitigates overfitting to demonstration trajectories. All auxiliary components are used only during training, allowing direct reuse of pretrained VLA weights and preserving the original inference architecture.
Should the Action Expert Train the VLM? Action-Token Supervision and Gradient Routing in Mixture-of-Transformer VLAs
Vision-language-action (VLA) models increasingly pair a pre-trained vision-language model (VLM) with a flow-matching expert that conditions its action predictions on the VLM's key-value representations. Those representations must carry information useful for control, yet the VLM was never trained for the action expert's flow-matching objective. Two signals can address this: gradients backpropagated from the expert's loss (coupling), or VLM supervision to predict discrete tokens encoding the same target actions. Prominent training recipes forgo the first, blocking expert-to-VLM gradients (insulation) to protect the VLM's existing capabilities, while also supervising it on vision-language data and action tokens. Because vision-language co-training also aims to preserve those capabilities, the additional protection provided by insulation is unclear. At the same time, insulation removes a direct signal about which features continuous action prediction requires. On the other hand, action-token supervision offers an alternative way to learn action-relevant representations, which may explain why insulation succeeds in practice. We disentangle these factors in autonomous driving and manipulation by independently varying action-token supervision and whether the expert's loss updates the VLM. In open-loop driving, insulation is competitive with coupling when action tokens are supervised; without them, coupling reduces trajectory-prediction error by roughly 25%. When both components are trained jointly from the outset, coupling increases closed-loop driving route progress by approximately 17% with action-token supervision and 40% without it, and similarly improves robot manipulation performance under both settings. These findings show that the value of gradient insulation depends on the VLM's action supervision and the evaluation setting, challenging insulation as a default VLA training recipe.
AceAction-5M: Hierarchical Goals for Generalist Action Learning Across 3D Game Worlds
Video games provide a practical setting for collecting vision-language-action (VLA) data, whose acquisition in physical environments is costly. Existing datasets, however, remain limited by restricted game coverage, sparse language supervision, or temporal hierarchies that do not necessarily reflect changes in task objectives. We introduce AceAction, a dataset of human gameplay across multiple 3D games for learning skill composition and subgoal transitions. AceAction organizes demonstrations around complete tasks, linking overall instructions, ordered subgoals, and local skill instructions to temporally aligned visual observations and keyboard-and-mouse actions. To construct these annotations, we develop a vision-language model pipeline that identifies subgoals from changes in objectives or interaction targets, anchors navigation goals to visually recognizable waypoints, and composes adjacent subgoals into task instructions. The resulting demonstrations cover navigation, reconnaissance, combat, interaction, and recovery, with explicit stage boundaries that connect task progression to fine-grained control. We assess annotation quality through model-based verification and human inspection. An audit of 500 trajectories yields a 95.8% pass rate across independent reviewer assessments. AceAction provides a data foundation for studying how agents execute, combine, and switch skills to follow complex instructions in dynamic game environments.
Vision-Language Models Are Joint Denoisers in Robotic Action Generation
Diffusion-based vision-language-action (VLA) models cascade pretrained vision-language models (VLMs) with diffusion action experts, but typically restrict the VLM to providing static conditioning. This separation limits how much pretrained perception and reasoning can guide action generation, especially in out-of-distribution settings. We introduce JADE, a Joint Action DEnoiser that enables the VLM and action expert to jointly refine actions. Our key insight is that structured action latents allow the VLM to read the current noisy action together with visual observations and language instructions. At every denoising step, the VLM contextualizes the noisy action and guides the action expert through gated cross-attention, so that its guidance evolves with the action being generated. We further supervise the VLM to recover clean action latents from noisy ones, which gives it a direct action-learning signal and couples perception with control. Additionally, joint denoising enables JADE to generate actions with fewer sampling steps, helping offset the additional cost of VLM participation. An adaptive step-size schedule further balances task performance and inference latency. On RoboDojo, JADE achieves an average success rate of 12.67%, the highest among the evaluated -based methods, and is competitive with recent VLA and WAM foundation models pretrained at scale. It also improves performance on RoboTwin 2.0 and raises the average success rate across four real-world bimanual tasks by 20.8 percentage points compared with .
BURGER: Benchmarking and Understanding Robotic Generalization via Embodied Drawing
How effectively can vision-language-action (VLA) models learn diverse language-conditioned behaviors as training data scales? Investigating this question requires large and varied robot datasets, whose collection on physical robots remains costly. We introduce BURGER, a language-conditioned robotic drawing benchmark with a scalable, prompt-driven pipeline for generating continuous drawing trajectories, both in simulation and the real world. In BURGER, an agent starts from the same blank canvas and controls a robot arm to draw the object or spatial composition specified by language, while each instruction allows an open-ended variety of valid drawings and stroke orders. Using 28,200 simulated demonstrations spanning 282 drawing tasks, we evaluate five representative VLA models across data scales, language grounding, and multimodal behavior. BURGER supports fine-grained analysis of language grounding, spatial composition, and action generation, revealing distinct failure patterns across VLA models. Our experiments suggest that some VLA models can benefit from greater task diversity and demonstration scale, capture multimodal drawing behaviors, and generalize to unseen compositions of familiar objects. Finally, we deploy a trained policy on a Franka Panda arm, demonstrating the ability to draw reactively under human intervention. Project homepage and code can be found at https://siyingzhao.github.io/BURGER/.
PRIME W0: Physical Priors from World Models for Embodied Manipulation
Vision-Language-Action (VLA) models excel at instruction following, yet most remain feed-forward: they condition actions only on the current observation and discard both long-horizon temporal context and the hierarchical structure of VLM representations. We present PRIME W0, a hierarchical conditional flow-matching framework that closes this gap through three synergistic designs: (1)per-layer KV injection from a pretrained VLM into an Adaptive MoE action expert; (2)dynamics-aware conditioning via asynchronous world-model subgoal-image prediction and a DINOv2 historical-frame semantic bridge; and (3)Action Dynamic Manifold (ADM) flow matching with covariance-structured noise that respects robot kinematic coupling. PRIME W0 is trained in four stages---VLM pre-training, DAT/NTP visuomotor pre-training, DiT flow-matching decoder assembly with WM injection, and end-to-end SFT---and achieves state-of-the-art results on LIBERO and real-robot benchmarks, with systematic ablations validating each architectural component.
PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
Vision--Language--Action (VLA) models generate robot actions from visual observations and language instructions, but their understanding of physical dynamics remains limited. Video world models can capture scene dynamics, yet generating future videos introduces high latency into the control loop. We introduce PDRVLA, whose lightweight Physical Dynamics Module extracts physical tokens from vision--language features to condition action generation without online video generation. The module learns through direct prediction of future RGB, depth, and optical-flow features, with auxiliary supervision through a frozen pretrained video world model. Training relies entirely on publicly available data, expanded through human-to-robot synthesis, subtask annotation, and embodied chain-of-thought (ECoT) annotation. Across four simulation benchmarks, PDRVLA achieves success rates of 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin~2.0, and 81.5% on SimplerEnv. Ablations show improved future prediction with world-model supervision and evaluate how subtask conditioning, future object-pose supervision, and ECoT training affect action success.
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
Evaluation-Guided Demonstration Refinement for Vision-Language-Action Policies: The Winning Entry in the MetaFine Open Competition
We fine-tuned a separate policy for each of the five tasks in the MetaFine Open Competition, with a task-specific training mixture and number of actions executed per prediction. Training started from the official demonstrations and expert procedures. Local evaluation exposed recurring grasping, pushing, and insertion failures. We addressed them by revising the demonstrated motions and by adding expert continuations from states the policy reached, while keeping nominal demonstrations in training. Checkpoints were compared on every instruction, so that a gain on one instruction could not hide a loss on another. Our submission, MAI, ranked first overall with a score of 96.66.
RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions
Scaling vision-language-action (VLA) models is constrained by scarce action-labeled demonstrations and incompatible control spaces across robot embodiments. We present RynnVLA-Latent, a framework that turns heterogeneous interaction videos into scalable supervision for robot policy pretraining. At its core, RynnLAM learns continuous, 3D-grounded latent actions from visual transitions without requiring native robot controls. By combining residual future-feature prediction, camera-compensated 3D-flow supervision, and camera-motion decoupling, RynnLAM encourages representations that capture control-relevant motion rather than appearance changes alone. We apply RynnLAM to 30.1K hours of human, robot, and simulation videos, producing 745.52M latent-action descriptors. RynnVLA-Latent then learns to predict latent-action chunks from visual observations and language through flow matching, before adapting its pretrained backbone and action expert to embodiment-specific robot actions. The resulting 4B policy achieves 98.8% success on LIBERO, 84.7% on LIBERO-Plus without benchmark-specific adaptation, 51.9% average success across five VLABench tracks, and 88.6% average success on the RoboTwin 2.0 bimanual benchmark. It also achieves the highest observed success rates among the compared methods on six real-world manipulation tasks across dexterous-hand and mobile-gripper platforms. Together with representation ablations and data-scaling results, these findings demonstrate the potential of latent-action supervision for scaling robot policy learning beyond native action-labeled data.
HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience
Extending Vision-Language-Action models (VLAs) to humanoid robots requires coordinated whole-body control and diverse demonstrations that are costly to collect in real-world scenarios. Although egocentric human demonstrations provide a complementary source of task experience, differences in visual embodiment and missing robot proprioception complicate transfer to humanoid policies. To address these challenges, we introduce HuGe, a family of Humanoid Generalist VLAs pre-trained on over 2,000 hours of egocentric human and robot data. Specifically, HuGe employs three existing representative VLA architectures to predict latent motion tokens for a pretrained whole-body controller, together with hand-control signals. Based on this, we introduce a three-stage training pipeline that combines egocentric pre-training, humanoid adaptation, and downstream fine-tuning. To facilitate transfer across embodiments, we construct pseudo-proprioceptive states through simulation playback during pre-training and modify arm appearance in egocentric observations through re-rendering during downstream fine-tuning. With this pipeline, HuGe achieves the state-of-the-art on simulation benchmark HumanoidArena. Furthermore, real-world experiments on the Unitree G1 indicate that augmenting robot demonstrations with visually adapted egocentric demonstrations significantly increases success on a pick-and-drop task.
TTT in VLA Models: Exploring Training with Human Demonstrations
Long-horizon robot tasks require retaining and using past information. We integrate Test-Time Training (TTT), which stores context in fast weights, into InternVLA-A1 and Pi0.5 to study action generation conditioned on human demonstrations through training on paired demonstrations and robot trajectories. We implement memory transfer between the two phases and truncated backpropagation, and examine training objectives, gradient coverage, and contrastive regularization across video, short-video, and single-frame cue tasks. In our experiments, adding TTT and training on paired data at the current scale did not lead to consistent target selection based on demonstrations. Under the tested settings, shorter sequences, full trajectory gradient connectivity, and contrastive regularization did not clearly improve responsiveness to demonstrations. When demonstrations were reduced to single-frame cues, real-robot execution still showed a preference for cups at fixed positions. We document the implementation, training configurations, and observations to support further inspection and replication.
DRILL: Training World Models to Improve Policies, Not Predict Pixels
World models are increasingly used to reduce environment interaction when post-training vision-language-action (VLA) policies with reinforcement learning. Yet they are typically trained to predict pixels one step ahead, even though their role here is not merely to render visually faithful videos but to produce learning signals from imagined rollouts that improve the policy. We show that this mismatch is substantial: lower pixel MSE can produce worse policy gradients, and a world model trained for policy utility can yield better downstream policies despite 17% higher pixel MSE on the next frame. We propose the Downstream-Return Imagination Learning Loop (DRILL), a bilevel framework that updates the world model to maximize the policy's return after an inner GRPO update on imagined rollouts. DRILL has two instantiations. DRILL-VWI is a closed-form surrogate, requiring no additional simulator rollouts, that reweights prediction errors by policy visitation, the group-relative GRPO advantage, and model confidence. DRILL-IMG is a full meta-gradient method that differentiates through the inner update using truncated Hessian-vector products. We show that DRILL-VWI recovers the local one-step component of the DRILL-IMG meta-gradient under score-compatibility and small-residual assumptions. Across five manipulation simulators and two VLA backbones, DRILL-IMG raises mean success rate by 13.1 points over WMPO and 7.9 points over RLVR-World, and complements RLVR pretraining rather than competing with it. More broadly, we believe video world models for VLA reinforcement learning can be productively trained by the policy gradients they produce, not only by pixel-level prediction.
AQUAACT: A Joystick-Level Underwater Teleoperation Dataset and Closed-Loop Benchmark
Underwater robots support a growing range of work, from inspecting offshore structures and pipelines to surveying reefs and accompanying divers. Most of these tasks, however, are still flown by hand by a pilot, which raises the question of whether they can be carried out autonomously instead. On land, vision-language-action (VLA) models already let robots carry out tasks autonomously from language instructions, which makes them a natural candidate for taking the pilot's place. Underwater, however, there is little data to train such models. We take inspiration from how these tasks are flown today: the pilot watches the camera image and steers with a joystick, whose commands the flight controller turns into motion. We therefore record the pilots and introduce AQUAACT, to our knowledge the first publicly available language-annotated underwater dataset pairing camera images with human joystick commands. It comprises 1.9 hours of real-world and 17.9 hours of simulated demonstrations, both flown by human pilots, together with the vehicle's inertial measurements. Experiments with AQUAACT show that its simulated demonstrations transfer to real data and greatly reduce the real demonstrations a policy needs. A closed-loop benchmark of seven policies in simulation, from imitation-learning baselines to VLA models, each trained with its default recipe, shows that these tasks remain far from solved: no policy completes more than a fifth of the episodes. Pretraining on land data gives large VLA models no advantage over small policies trained from scratch, a gap that land-trained models have yet to close underwater.
VLA-: Learning Physical State Transition for Vision-Language-Action Models
Vision-language-action (VLA) models transfer semantic knowledge acquired through vision-language pretraining to robot control, achieving strong performance across various benchmarks. However, most VLAs directly map observations to demonstrated actions without explicit supervision of intermediate physical state changes, making them prone to overfitting observation-action correspondences. Some methods introduce intermediate reasoning or future-state prediction in language, visual, or latent spaces to guide action generation, but do not explicitly supervise the physical state transition itself. We therefore propose VLA-, a general framework that uses physical changes between current and future states as explicit learning targets. To support this objective, we construct a large-scale pretraining dataset and derive physical state-transition annotations covering task-relevant semantics and visual motion changes, providing more direct supervision for continuous action generation. During inference, the model generates actions directly from current observations without additional inference overhead. VLA- achieves performance competitive with the state of the art on LIBERO and RoboTwin, and outperforms in real-robot experiments. Our code is available at https://github.com/ZJU-OmniAI/VLA-Delta
Do VLAs Really Follow Instructions? Diagnosing How-Level Instruction Following in Vision-Language-Action Models
Vision-language-action (VLA) models are expected to follow instructions like their language backbones, yet prior work has focused on What-level instructions that specify the task rather than How-level instructions that specify how to carry it out. Through our diagnoses, we find that How-level instructions do not help current VLAs and lower task success. We hypothesize that pairing each training episode with a single instruction limits How-level instruction following, and show that training with segment-level enriched instructions makes actions respond to such instructions without consistently raising task success. To measure How-level instruction following beyond standard benchmarks, we introduce SIMPLER-IN, where success requires conditions given only in the instruction. It shows that the gained responsiveness does not extend to demanding or out-of-distribution How-level instructions. How-level instruction following thus remains an open challenge.
What Matters in Quality-Aware VLA Pretraining: A Benchmark and Empirical Study
How should demonstration quality be used during vision-language-action (VLA) pretraining to improve performance after target-task fine-tuning? We compare weighting, filtering, and conditioning across reward definitions and label granularities using 5,000 naturally suboptimal human demonstrations and 24 target tasks. Within each VLA, we match demonstrations and training budgets. Additional comparisons examine pretraining data size and the roles of quality tags during fine-tuning and inference. Across three VLAs, we find that approach-aware episode conditioning improves mean success over plain behavioral cloning; conditioning and filtering favor episode and chunk labels, respectively, on average, with model-dependent reversals; and constant-tag fine-tuning improves mean success after conditioned pretraining with either tagged or untagged inference. We introduce AXIS-Bench to support these controlled comparisons. Built on crowdsourced simulation demonstrations, it provides recorded object and contact states for recomputing quality labels on shared trajectories. Replayable trajectories support workspace-aligned rendering and user-defined real-world target tasks. In our real-world experiment, simulation pretraining raises success from 9.6% to 31.7%. Project page: https://axis-bench.github.io/.
Axiom-0: Hierarchical Self-Refinement for Vision-Language-Action Policies
Reliable manipulation requires policies that generalize across tasks, layouts, and object interactions. Vision-language-action (VLA) models combine pretrained multimodal understanding with learned control, but their performance remains limited on unfamiliar task compositions. We present Axiom-0, a VLA policy that addresses both the quality of its representations and their use in action generation through hierarchical self-refinement. Intra-VLM Representation Refinement (IVRR) strengthens selected intermediate features through consistency supervision with learned feature projections during training. VLM-Action Interface Refinement (VAIR) recursively updates the resulting conditioning before action generation. Joint optimization of the supervised VLM layers, VAIR, and the action generator connects the two components. On RoboCasa365, Axiom-0 achieves 62.32% overall success in our evaluation, exceeding the highest published leaderboard score (Paimon-0, 58.1%) by 4.22 percentage points and the Xiaomi-Robotics-1 reference by 4.92 points. Under the evaluated profiling setup, mean action-chunk generation latency increases by 1.7% over Xiaomi-Robotics-1. These results demonstrate improved task completion with modest inference overhead.
Same Scene, Different Task: Skill Alignment for Compositional Generalization in VLAs
Vision-language-action (VLA) models often struggle to generalize to skill combinations absent from their fine-tuning demonstrations, even when every constituent skill has been demonstrated. We focus on a vision shortcut as one failure mode: during fine-tuning, visual observations can serve as a proxy for the instruction, so a policy may execute a demonstrated combination associated with similar observations rather than the instructed combination. This motivates training with counterfactual pairs formed by holding a demonstration observation fixed while changing the instruction to specify an undemonstrated combination. These pairs, however, lack corresponding demonstrated action targets. Crucially, the currently required skill has already been demonstrated, but actions from those executions cannot serve as direct targets because the same skill can require different actions across observations. We propose CRAFT, which transfers supervision from demonstrated executions of the required skill to counterfactual pairs using skill representations that can be reused across executions of the same skill. Across three VLA models and two simulation benchmarks, CRAFT improves success on undemonstrated combinations while maintaining high success on demonstrated ones; it also improves compositional generalization on a real robot. Project website: https://taegeunyang.github.io/craft/
DistAL: Distance-based Advantage Learning for VLA Fine-Tuning
Vision-language-action models (VLAs) have trans- formed the field of robotic manipulation in recent years by combining the semantic understanding of LLMs with the precise control of flow-matching policies. Advantage conditioning is a recent technique that iteratively improves VLAs by training a value function on deployment data and using this to train an advantage-conditioned policy. Previous works have only applied simple, low-information success/failure rewards, which leave the value function unable to distinguish states of differing quality beyond how far along the task they appear. Motivated by an exploration of out-of-distribution (OOD) detection methods, we introduce Distance-based Advantage Learning (DistAL), which, by using an embedding space distance as a reward, produces a more informative value function and subsequently a higher downstream task success rate. We validate our method on a series of simulation benchmarks and dexterous bi-manual manipulation tasks on real hardware.
Decoupling Vision, Language, and Action for Efficient Multi-Task Robot Policies
Vision-Language-Action (VLA) policies commonly run Vision-Language Model (VLM) backbones with billions of parameters at every policy inference, which costs latency and energy. We revisit a decoupled alternative for multi-task manipulation: separate vision and language encoders whose representations condition a compact action head. We run a standardized comparison that varies the vision encoder, the language encoder, and the action head while holding the demonstrations, the training-step budget, the tasks, the evaluation protocol, and the measurement platform fixed, against seven VLA baselines. The resulting Decoupled Embodiment Model (DEM) combines a fine-tuned DINOv3 vision encoder, a frozen NeoBERT language encoder, and a MeanFlow head that generates an action chunk in one forward pass. On 18 RoboCasa tasks evaluated with held-out instruction paraphrases and randomized scenes, DEM reaches 55.6% mean success against 56.9% for GR00T N1.7 and 54.6% for , and on three real-robot tasks it reaches 66.0% against 68.0% for GR00T N1.7. On the same workstation, DEM needs 6.1,ms per policy forward pass, a maximum throughput of 162.7 policy calls per second, and draws an estimated 2.07,J of GPU energy per call, eight to seventeen times the throughput and six to fifteen times less energy than these VLM-backbone policies. Within this trained-task regime, DEM sits on the observed success--latency--energy frontier and provides a strong, efficient baseline for language-conditioned robot skills.
IMLE-VLA: Fast Single-Step Action Generation for Vision-Language-Action Policies
Vision-language-action (VLA) policies leverage pretrained vision-language backbones to achieve strong cross-task generalization. A leading design couples this backbone with a dedicated continuous action head trained via diffusion or flow matching. However, such heads rely on iterative multi-step sampling, for example 10 Euler steps in . This creates an inference bottleneck that produces stop-and-go movement in the robot and slower task completion. We introduce IMLE-VLA, which replaces the iterative action head with a single-step conditional generator trained via conditional Implicit Maximum Likelihood Estimation (cIMLE). The cIMLE objective promotes multimodal action coverage, avoiding the mode collapse of naive regression heads while eliminating multi-step sampling entirely. When IMLE-VLA is applied to , it increases inference frequency 3.67x (55 Hz vs. 15 Hz), enabling up to 11x higher action throughput. On the 40-task LIBERO benchmark, IMLE-VLA achieves the highest average success rate (98.0%) among all baselines while leading in inference frequency. Under the test-time perturbations of LIBERO-plus, IMLE-VLA retains 's robustness while other baselines degrade sharply, confirming that the cIMLE head preserves generalization. Real-world experiments on a Franka Emika Panda across four tasks demonstrate smoother motion (2.2x to 3.0x lower jerk) and faster task completion, with IMLE-VLA outperforming on every task and reducing average VLA inference time per episode by 3.9x to 6.6x. Videos and code are available at https://kianhk6.github.io/IMLE-VLA/
HuRo: Robotizing Human Videos for Scalable VLA Pretraining
Human video datasets offer an abundant and diverse source of interaction data that can complement expensive real-robot data. To bridge the human-to-robot embodiment gap, existing approaches either robotize videos in task-matched settings or address observation and action alignment separately. In this work, we systematically examine whether robotized human videos can serve as an effective and scalable source of robot-aligned supervision. To this end, we develop a robotization pipeline that converts heterogeneous human videos into robot-aligned observations and action trajectories while inferring missing intermediate signals across annotation levels. Using this pipeline, we construct the HuRo dataset, comprising about 630K robotized episodes and 142M processed frames from five human-video sources. Across four real-world manipulation tasks, pretraining a VLA policy on increasing amounts of robotized human-video data improves overall completion from 51.5% to 80.3% and OOD completion under spatial and visual shifts from 34.9% to 72.2%. Ablations further show that visual robotization improves OOD robustness and that end-to-end pretraining with retargeted actions outperforms visual-only transfer. Project website: https://3587jjh.github.io/HuRo.
ICI-VLA: In-Context Imitation with Spatiotemporally Aligned Demonstrations for Vision-Language-Action Models
Vision-Language-Action (VLA) policies are commonly adapted to new manipulation settings through additional gradient updates, which limits rapid deployment when task-specific data or compute is scarce. We present ICI-VLA, a training and retrieval framework that equips a text-action VLM with few-shot test-time adaptation through in-context demonstrations. Unlike mainstream VLA designs based on action-specific multimodal fusion, ICI-VLA retains the native text-generation interface. ICI-VLA updates its parameters only during offline training; at inference, the policy remains fixed and conditions action generation on retrieved micro-demonstrations. The framework decomposes long trajectories into short, semantically labeled examples and trains an RD-Encoder with positives mined by Dynamic Time Warping (DTW), aligning the retrieved context with the phase and geometry of the current subtask. We further introduce Target Action Masking, a context-corruption objective designed to reduce direct action copying and increase reliance on the current observation. ICI-VLA reaches average success rates of 97.7% on LIBERO and 60.4% on RoboTwin 2.0, exceeding the highest reported baseline average on RoboTwin 2.0 by 19.3 percentage points. It also achieves 83.2% across four physical tasks. These results indicate that a fixed VLA policy can benefit from conditioning on spatiotemporally aligned demonstrations at test time.
Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models
Vision-Language-Action (VLA) models can turn multimodal context into robot actions, but their action decoders are still trained largely by behavior cloning. This supervises which motor command was demonstrated while leaving implicit the local objective served by the behavior under the instruction. Future-based supervision enriches action learning with frames, latent observations, trajectories, or motion representations, but these signals capture particular realizations of what may happen rather than the shared semantic objective of the forthcoming behavior. We propose Intention Distillation (INDI), which distills behavior-level intent into the action decoder. During training, a frozen teacher VLM interprets a demonstrated segment from the current observation, instruction, coarse action summary, and corresponding execution video. From its standard inputs, the deployed VLA recovers the resulting multimodal intent representation at an intermediate decoder layer and uses it to organize action prediction together with representations of how the behavior unfolds and what it achieves. On SimplerEnv-Bridge, INDI improves GR00T-N1.7 from 64.3% to 84.7%, and on RoboCasa Kitchen it improves the controlled GR00T-N1.7 baseline from 64.1% to 70.3%, with consistent gains on across both benchmarks. In real-world tasks, INDI improves average success from 62.0% to 68.7%, with gains of up to 12.0 pp on longer-horizon tasks. Further analyses show that the recovered latent is used by the decoder, captures behavior objective and execution progress, and organizes downstream predictions in an objective-dependent manner. These results show that action decoders benefit from explicitly modeling the semantic objective of the behavior they generate.