PaperlayerView exact revision on arXiv

Source-linked AI summary

AR-VLA: True Autoregressive Action Expert for Vision-Language-Action Models

Yutong Hu, Jan-Nico Zaech, Nikolay Nikolov, Yuanqi Yao, Sombit Dey, Giuliano Albanese, Renaud Detry, Luc Van Gool, Danda Paudel

arXiv:2603.10126v2cs.ROcs.AI

TL;DR

Reactive VLA and diffusion policies reset temporal context at each observation, limiting persistent history for streaming robot control. AR-VLA uses a stateful autoregressive action expert with refreshable vision-language context, achieving the highest average success rate of 61.5% and smoother trajectories than reactive baselines.

  • Problem

    Reactive VLA and diffusion policies generate isolated action chunks without persistent perception-action history, limiting temporal awareness for streaming manipulation control.

  • Method

    AR-VLA uses a standalone autoregressive action expert with a Hybrid Key-Value Cache and Dynamic Temporal Re-anchoring to combine rolling action history with refreshable vision-language context.

  • Results

    61.5% average success rate made AR-VLA the highest-performing policy, outperforming CogACT by a +9.4% margin (52.1%) across four WidowX manipulation tasks.

  • Takeaways & Limitations

    AR-VLA provides a scalable action-generation framework that structurally decouples motion syntax from semantic perception while supporting smoother, more temporally consistent control.

  • Takeaways & Limitations

    AR-VLA is sensitive to out-of-distribution trajectories because errors in its cached action history can trigger compounding feedback loops during rollouts.

Abstract

from arXiv · show

We propose a standalone autoregressive (AR) Action Expert that generates actions as a continuous causal sequence while conditioning on refreshable vision-language prefixes. In contrast to existing Vision-Language-Action (VLA) models and diffusion policies that reset temporal context with each new observation and predict actions reactively, our Action Expert maintains its own history through a long-lived memory and is inherently context-aware. This structure addresses the frequency mismatch between fast control and slow reasoning, enabling efficient independent pretraining of kinematic syntax and modular integration with heavy perception backbones, naturally ensuring spatio-temporally consistent action generation across frames. To synchronize these asynchronous hybrid V-L-A modalities, we utilize a re-anchoring mechanism that mathematically accounts for perception staleness during both training and inference. Experiments on simulated and real-robot manipulation tasks demonstrate that the proposed method can effectively replace traditional chunk-based action heads for both specialist and generalist policies. AR-VLA exhibits superior history awareness and substantially smoother action trajectories while maintaining or exceeding the task success rates of state-of-the-art reactive VLAs. Overall, our work introduces a scalable, context-aware action generation schema that provides a robust structural foundation for training effective robotic policies. Code and Videos available at https://arvla.insait.ai

I. INTRODUCTION · II. RELATED WORK · A. Vision-Language-Action models (VLAs).

The paper argues that existing VLAs and diffusion policies are reactive despite being called autoregressive, because they generate isolated action chunks rather than continuous, history-aware control. AR-VLA addresses this by introducing an independent autoregressive Action Expert that maintains persistent action context while remaining decoupled from vision-language perception.

  • I. INTRODUCTION: Existing VLAs labeled “autoregressive” are deceptive because they do not model robot actions as a continuous causal stream.The paper contrasts action generation with the causal sequence modeling used in language and vision-language models.
  • I. INTRODUCTION: Diffusion Policies and existing VLAs typically use action chunking, predicting static action blocks at each perception step and therefore reacting to local snapshots.These methods can provide short-horizon smoothness but structurally reset temporal context between observations.
  • I. INTRODUCTION: AR-VLA achieves competitive or superior benchmark performance and significantly smoother, more kinematically consistent trajectories than reactive policies, including OpenVLA, Flow-Matching, ACT, and Diffusion Policy.The figure reports results for both generalist and specialist benchmarks and visualizes joint trajectories over time.
  • I. INTRODUCTION: Effective manipulation requires both situational awareness of the workspace and temporal awareness of prior motion and end-effector dynamics.The paper argues that VLMs provide situational understanding but are structurally ill-suited to temporal control awareness.
  • I. INTRODUCTION: A true Autoregressive Action Expert captures causal dependencies across the trajectory, supports high-frequency motor execution despite perception latency, and separates action generation from the VLM backbone.These properties are presented as benefits of treating action as a “language of motion.”
  • I. INTRODUCTION: AR-VLA instantiates an independent Action Expert that decouples high-level vision-language reasoning from high-frequency temporal control for both specialist and generalist robot policies.The architecture maintains a continuous evolving state rather than treating the action head as a dependent appendage of the VLM.
  • I. INTRODUCTION: The core contribution formulates action generation as causal sequence modeling across time using a long-lived context of past actions.The paper identifies temporal inconsistency and jitter as problems of reactive policies and introduces the Hybrid Key-Value Cache as one technical pillar.
  • A. Vision-Language-Action models (VLAs).: AR-VLA targets generalist-policy limitations by integrating persistent historical context into the autoregressive process without changing the Vision-Language perception component.This builds on work that injects internet-scale VLM priors into low-level action generation to broaden beyond task-specific imitation-learning data.

B. Action Representation and Pretraining. … A. Problem Formulation

The paper motivates learned action representations and persistent memory for long-horizon robotic control, then formalizes AR-VLA as a stateful actor that preserves continuous kinematic history while conditioning on refreshed visual-language context. This design addresses the temporal incoherence and latency limitations of reactive, stateless action prediction.

  • B. Action Representation and Pretraining.: Modern action representation learning seeks well-structured spaces for cross-modality alignment beyond normalization, categorical binning, and k-means discretization.These classical methods are emphasized as traditional solutions for the relatively low-dimensional action domain.
  • C. Architectures with context awareness: Long-horizon, non-Markovian robotic applications require memory mechanisms that use historical observations and actions.Recurrent and Transformer-based policies established memory as important in partially observable environments.
  • III. METHODOLOGY: AR-VLA uses a stateful two-stage architecture to bridge high-latency semantic perception and high-frequency motor control.Its standalone autoregressive action expert maintains kinematic continuity with a Hybrid Key-Value Cache while combining proprioceptive history and refreshed visual-linguistic context.
  • A. Problem Formulation: A robot trajectory is formalized as a sequence of observations and actions, with observations decomposed into visual frames, language instructions, and proprioceptive states.The formulation represents trajectories as {(o_t, a_t)^T} and separates exteroceptive inputs from proprioception.
  • A. Problem Formulation: Temporal observation stacking and chunked prediction create pseudo-history but remain Markovian, forcing repeated history and velocity inference that can cause jittery, temporally incoherent control.The limitation arises because each new observation window requires the model to re-infer historical intent and current velocity.
  • A. Problem Formulation: The Reactive Actor maps the current observation to the current action while resetting context memory at every step.Its observation embedding is produced by a perception encoder Φ(·), such as a VLM or smaller backbone.
  • A. Problem Formulation: The AR Actor predicts each action from continuous causal kinematic history while conditioning on the most recently processed visual-language prefix.With i ≤ t denoting the latest processed visual frame, the persistent chain a<t, s<t supports kinematic smoothness and robustness to visual-language latency.

B. Model Structure

AR-VLA uses a unified Transformer decoder with continuous action tokens and a Hybrid Key-Value Cache that separates long-lived proprioceptive history from refreshable vision-language context. Dynamic Temporal Re-anchoring uses RoPE-based relative timing to represent visual-context staleness consistently across training and inference.

  • Unified Decoder with Hybrid Cache: The architecture asynchronously bridges a VLM backbone and autoregressive Action Expert by injecting temporally re-anchored vision-language tokens into rolling kinematic history.The re-anchored VL tokens form a semantic prefix, while incrementally advancing step embeddings query the shared cache to generate future action sequences.
  • Continuous Action Representation: Continuous actions are projected into the model dimension as one token per timestep, with outputs regressed to the next action.Actions represent end-effector pose deltas or joint velocities and preserve continuous regression precision for low-level manipulation.
  • Unified Decoder with Hybrid Cache: The Hybrid Key-Value Cache maintains separate context streams, including a rolling FIFO buffer for robot state and action history.This long-lived proprioceptive window is designed to capture momentum beyond the 1–4-step history stacks used by reactive VLAs.
  • Dynamic Temporal Reanchoring: Dynamic Temporal Re-anchoring uses RoPE to encode relative distance between the high-frequency action stream and atemporal vision-language context.This mechanism addresses temporal alignment when perception and control operate at different frequencies.
  • Dynamic Temporal Reanchoring: Action tokens follow executed timesteps, whereas VLM tokens receive manually assigned anchor indices so attention reflects their temporal validity.The resulting relative distance m − n represents visual-context staleness, while global time shifts preserve the same relative-distance attention behavior.
  • Dynamic Temporal Reanchoring: A training example with m = 25 and n = 20 yields ∆t = 5, matching inference at global step 500 with a VL update from step 495.Thus, the same visual grounding logic applies at steps 25 and 500 when staleness remains ∆t = 5.

C. Training Details

Training uses two phases: standalone action-only pretraining first masters motion syntax, then VL-action alignment connects the expert to perception through DTR and anchored supervision. This separates kinematic skill learning from VL grounding.

  • Phase 1: Action-Only Pretraining: The protocol first pretrains the action generator on large-scale trajectories as a standalone autoregressive model.This action-only phase is designed to master motion syntax before perception grounding.
  • Phase 1: Action-Only Pretraining: Causal masking and sequential RoPE indices optimize the action sequence modeling objective.
  • Phase 1: Action-Only Pretraining: The resulting proprioceptive expert independently learns kinematic syntax, including joint limits, motion profiles, and common move patterns.
  • Phase 2: VL-Action Alignment: The second phase connects the VLM backbone to the action expert using DTR for VL-action alignment.Given (x_past, v, l, x_fut), the protocol uses priming, anchoring, and stochastic supervision with historical dropout.
  • Phase 2: VL-Action Alignment: During priming, history x_past enters the actor with indices {0, . . . , H−1}, while VL features from (v, l) receive fixed index H at the history-future junction.

D. Inference Details

At inference, AR-VLA autoregressively predicts actions from persistent action history and refreshable visual-language prefixes, remaining effective when perception is temporally outdated. Its decoupled dual-thread design supports asynchronous control and perception, enabling lower latency and consistent control frequency.

  • Conditional autoregressive inference: AR-VLA predicts precise actions conditioned on an outdated visual-language prefix through DTR and teacher-forcing training.DTR makes attention generalize across temporal offsets Δt = t − i.
  • Hybrid KV cache: Inference uses a hybrid KV cache combining persistent action history with the most recently available visual-language tokens.The cache is constructed and maintained dynamically during runtime.
  • Hybrid KV cache: The action cache preserves long-term trajectory history, while the visual-language cache refreshes as new embeddings arrive and applies DTR to re-anchor timestamps.The visual-language cache functions as a single-slot FIFO snapshot.
  • Asynchronous deployment: A decoupled dual-thread architecture lets the high-frequency Action Thread and native-frequency Perception Thread execute asynchronously in serial or parallel modes.The Action Thread generates actions and updates the action cache, while the Perception Thread processes visual-language inputs.
  • Asynchronous deployment: Compared with denoise-based chunking, the lightweight action head lowers serial per-step latency and parallel execution maintains consistent control frequency without blocking on specific visual-language frames.Parallel prediction conditions on internal kinematic history and the latest valid visual-language prefix.

IV. EXPERIMENTS · A. Generalist and Specialist Policy Performance

The experiments evaluate AR-VLA as an action expert for specialist and generalist policies, emphasizing history awareness, high-frequency control, efficiency, trajectory quality, and long-horizon behavior. Across simulation and manipulation benchmarks, AR-VLA and AR Actor achieve strong performance, including leading generalist success and competitive specialist results.

  • IV. EXPERIMENTS: The evaluation tests whether AR-VLA can replace standard action heads in specialist and generalist policies while improving history awareness and high-frequency control.The study also examines inference efficiency, trajectory quality, long-horizon tasks, causal pretraining, RoPE anchoring, history masking, and autoregressive context length.
  • IV. EXPERIMENTS: The experiments span generalist and specialist policies across diverse embodiments, action spaces, and tasks, including history-dependent PushT2 and Stack3 scenarios.PushT2 and Stack3 are designed so critical information becomes unobservable without maintaining action history.
  • A. Generalist and Specialist Policy Performance: Generalist comparisons use BridgeV2-trained, identical-scale 3B + 300M models sharing a Paligemma-3B VLM backbone and knowledge insulation.The compared action experts include Fast-token decoding, chunk prediction, and AR-VLA-style autoregressive generation.
  • A. Generalist and Specialist Policy Performance: 61.5% average success rate: AR-VLA leads four WidowX SimplerEnv tasks, outperforming CogACT by +9.4% (52.1%).The tasks are put spoon on towel, put carrot on plate, stack green block on yellow block, and put eggplant in a yellow basket; AR-VLA shares the VLM backbone with Pi-0-Fast and Pi-0.5.
  • A. Generalist and Specialist Policy Performance: The results attribute AR-VLA’s consistent sequence generation to stored history key-value caches and suggest architectural simplicity supports task-agnostic performance.These observations are presented as aligned with prior ARP findings.
  • A. Generalist and Specialist Policy Performance: 97.33% scripted success and 67.33% human demonstration success: AR Actor outperforms ACT (86.0%/50.0%) and Diffusion Policy (33.33%/10.0%) on ALOHA cube transfer.On ALOHA peg insertion, AR Actor reaches 54.67% scripted success versus ACT’s 32.0%.
  • A. Generalist and Specialist Policy Performance: 65.20% success rate: Diffusion Policy leads PushT, while AR Actor remains competitive across specialist tasks.The reported pattern suggests Diffusion Policy excels on specific tasks but performs less consistently across tasks.

B. Efficiency and Smoothness Analysis · C. History-Awareness Evaluation

AR-VLA decouples control from perception to sustain fast action execution despite slower visual-language inference, while producing low-jerk trajectories. Its stored action-history context supports long-horizon tasks requiring memory of states that become visually unobservable, outperforming reactive policies on both benchmarks.

  • B. Efficiency and Smoothness Analysis: AR-VLA maintains a stable 29ms per-action control frequency while the perception backbone is capped at 70ms per frame.This stability comes from structurally decoupling the control thread from VLM inference.
  • B. Efficiency and Smoothness Analysis: AR-VLA achieves the lowest maximum and average jerk during task execution.The passage attributes this result to maintaining rigorous spatial and temporal act—.
  • C. History-Awareness Evaluation: PushT2 tests whether the policy remembers which of two goals has been reached after that information becomes unobservable midway through execution.The task requires pushing a T-shaped block to both goal positions in arbitrary order.
  • C. History-Awareness Evaluation: Both history-awareness tasks require memory of unobservable past states rather than relying only on the current observation.In PushT2, the visited goal becomes unobservable; in Stack3, the battery location becomes unobservable after being covered.
  • C. History-Awareness Evaluation: Stack3 tests whether the policy retrieves a battery’s location from action history after the battery becomes occluded beneath a cup.The real-world task then requires stacking another cup on top.
  • C. History-Awareness Evaluation: AR-VLA substantially outperforms existing policies on both history-awareness tasks, whereas reactive policies can oscillate between sub-goals after losing context of completed steps.The passage characterizes this failure mode as “temporal amnesia.”

D. Ablations on Design Decisions · V. CONCLUSION AND DISCUSSION

The ablations validate key AR-VLA design choices, including causal pretraining, masking, and extended action-history context. The discussion frames autoregressive action generation as a shift toward temporally consistent streaming control while noting OOD, knowledge-insulation, and visual-history challenges.

  • D. Ablations on Design Decisions: Removing Phase 1 pretraining doubles Phase 2 convergence time and produces inferior final performance.Without causal pretraining, the model must learn joint kinematics from scratch instead of building on pretrained motion priors.
  • D. Ablations on Design Decisions: A 0.6 mask rate provides the optimal balance between leveraging historical context and robustness to prediction errors.Lower robustness can arise when the model over-relies on its own history and predictions deviate during rollouts.
  • D. Ablations on Design Decisions: Increasing the action KV-cache length from 1 to 40 steps yields an upward trend in success rate, validating extended temporal context.The ablation directly links longer action history with improved performance.
  • V. CONCLUSION AND DISCUSSION: The Autoregressive Action Expert shifts control from reactive, snapshot-based predictions to continuous streaming action sequences.This design equips the policy with a persistent causal history of its own.
  • V. CONCLUSION AND DISCUSSION: Persistent causal history addresses temporal inconsistencies in current VLA architectures and keeps action generation consistent across spatial and temporal dimensions.The reported empirical benefits include significantly smoother trajectories and reduced execution latency versus reactive baselines.
  • V. CONCLUSION AND DISCUSSION: The cache-based autoregressive paradigm can compound individually familiar states into novel KV-cache sequences that cause out-of-distribution failures.This is identified as a new challenge introduced by the autoregressive design.
  • V. CONCLUSION AND DISCUSSION: Direct autoregressive action gradients can degrade the VLM’s pretrained semantic priors, requiring a knowledge-insulation training strategy.The discussion compares this challenge with flow-matching experts.
  • V. CONCLUSION AND DISCUSSION: Although proprioceptive history is fully cached, visual processing remains an unresolved limitation of the proposed paradigm.The passage identifies incomplete visual-history handling as a remaining challenge.

Appendix · A. Model Architecture

The appendix outlines implementation, benchmarking, history-awareness, execution, and limitation details, while the model architecture combines PaliGemma with a matched autoregressive Action Expert. The design preserves model-scale parity with baselines while replacing chunked or denoising objectives with next-action prediction.

  • Appendix: The appendix provides detailed configurations, hyperparameters, benchmark protocols, history-awareness tasks, execution snapshots, and discussion of limitations.These materials cover PaliGemma and Action Expert layers, Generalist and Specialist settings, scoring calibration, PushT2 and Stack3, experiments, and streaming-VLM/VLA issues.
  • A. Model Architecture: PaliGemma integrates a SigLIP-So400m visual encoder with a Gemma-2b language model comprising 27 visual layers and 18 language-decoder layers.The visual tower uses 14x14 patches, hidden dimension 1152, and feed-forward dimension 4304; the language model uses hidden dimension 2048 and feed-forward dimension 16,384.
  • A. Model Architecture: The autoregressive Action Expert has 18 transformer layers, matching Gemma-2b to align visual-linguistic semantics with kinematic sequences.Its structure is based on the design used by Pi-zero and is intended to support semantic-to-action alignment.
  • A. Model Architecture: The AR-VLA and π0.5-adapted BridgeV2 baseline share identical scale: a PaliGemma-3B backbone and a 300M-parameter transformer Action Expert.This establishes a controlled comparison between autoregressive streaming and the baseline training recipe.
  • A. Model Architecture: Knowledge insulation trains semantic reasoning and motor control through independent signals by blocking action-objective gradients from entering the VLM.The VLM is trained exclusively with an auxiliary Fast Token loss, while the action loss updates only the Action Expert.
  • A. Model Architecture: AR-VLA uses the π0.5 flow-matching Action Expert’s exact architecture and parameter budget but predicts actions autoregressively with a single-step next-action loss.The baseline generates action chunks through multi-round iterative denoising, whereas AR-VLA uses standard next-action prediction.
  • A. Model Architecture: Specialist evaluations compare AR-VLA with ACT and Diffusion Policy using shared ResNet-18 vision encoders and comparable implementation structures.The baselines use LeRobot implementations, and the AR actor is structured comparably to ACT while preserving parameter and architectural parity.
  • A. Model Architecture: The ACT adaptation repurposes its 4-layer encoder as a cached semantic-prefix perception backbone and its 4-layer decoder as an autoregressive Action Expert.This replaces the original action-chunk prediction design with a next-token prediction framework.

B. Hyperparameters

The models use cross-attending Action Experts with persistent internal history, built on different perception backbones for generalist and specialist settings. Generalist training separates kinematic-syntax learning from multimodal alignment, while temporal-consistency parameters and task-specific losses support their differing operating regimes.

  • Architecture: The generalist uses a FAST-insulated PaliGemma-3B backbone, whereas specialists use ResNet-18, with both Action Experts cross-attending to perception features and maintaining persistent history.The specialist configuration is based on a 4-4 Transformer-based Actor setup.
  • Training: Generalist training has two phases: action-only kinematic-syntax learning followed by multimodal alignment.Phase-1 uses a distinct action-only dataset rather than the BridgeV2 dataset used for final alignment.
  • Training: 20,000 steps with batch size 1024 typically complete Phase-1 in approximately 2 hours on one A6000 GPU.This phase is described as notably efficient.
  • Temporal consistency: Temporal-consistency settings use a larger history window for generalist long-horizon stability and higher operating frequencies for specialists’ precise bimanual tasks.The section identifies history and masking parameters as the controls used to maintain temporal consistency.
  • Losses: AR-VLA weights translation and rotation losses at λ = 1.0 and gripper loss at λ = 0.1, while specialists use standard L2 next-action loss without VAE latent modeling.The specialist objective focuses purely on autoregressive sequence dependency.

C. Benchmark Protocol and Completion Criteria · D. History Awareness Task: Design, Data Collection, and Execution

The benchmark protocol evaluates simulated and real-world tasks across varied initial conditions and repeated trials, while the history-awareness benchmarks test whether policies can retrieve hidden past states. PushT2 and Stack3 use distinct demonstrations and completion criteria, revealing failures of reactive policies and advantages of AR-VLA’s persistent autoregressive memory.

  • C. Benchmark Protocol and Completion Criteria: Each task varies M initial conditions and executes N trials per condition, reporting average task progress across tasks, configurations, and trials.This yields N × M trials per task under a protocol spanning simulated and real-world environments.
  • C. Benchmark Protocol and Completion Criteria: WidowX real-world episodes use one side-mounted external camera, with calibration requiring every model to achieve 3/3 on the “eggplant in sink” task before benchmarking.The setup mirrors BridgeV2 and uses the exact toy kitchen training configuration for calibration.
  • C. Benchmark Protocol and Completion Criteria: Real-world episodes time out after 200 steps, or 40 seconds at 5Hz, and receive the highest completion state reached unless manual termination is needed to prevent hardware damage.Collisions and unintended object displacements are otherwise permitted during evaluation.
  • D. History Awareness Task: Design, Data Collection, and Execution: PushT2 and Stack3 are designed so success requires remembering non-observable past states rather than relying only on immediate visual feedback.PushT2 is simulated, whereas Stack3 is performed on a real robot.
  • D. History Awareness Task: Design, Data Collection, and Execution: PushT2 requires placing a T-block at two goals in any order, while Stack3 requires covering a hidden battery with one cup and stacking a second cup on top.The occupied goal and battery location become unavailable from subsequent visual feedback.
  • D. History Awareness Task: Design, Data Collection, and Execution: The datasets contain 240 human-expert PushT2 episodes and 16 teleoperated Stack3 episodes collected across flexible orders and varying initial layouts.PushT2 demonstrations impose no target-reaching order, while Stack3 layouts vary cup and battery positions to preserve the hidden midway state.
  • D. History Awareness Task: Design, Data Collection, and Execution: PushT2 considers a goal reached when the T-block-to-goal Intersection over Union exceeds 0.9, while Stack3 scores progress from 0.25 to 1.00 across four manipulation stages.Stack3 stages are first-cup pickup, battery covering, correct second-cup pickup, and final stacking.
  • D. History Awareness Task: Design, Data Collection, and Execution: Reactive policies frequently fail through “temporal amnesia,” whereas AR-VLA’s autoregressive generation and persistent KV cache maintain task intent and kinematic momentum.The reported contrast concerns non-Markovian manipulation sequences, where reactive policies oscillate between sub-goals or fail to locate hidden objects.

E. Screenshot Highlights from Task Execution · F. Discussion and Limitations

The paper illustrates AR-VLA execution in simulated and real-world settings, including specialist AR-Actor behavior. It then discusses limitations involving error compounding, gradient insulation, modularity, and extending streaming principles to VLMs.

  • E. Screenshot Highlights from Task Execution: AR-VLA is illustrated through zero-shot task execution in the SIMPLER simulator.The section presents screenshots of simulated execution.
  • E. Screenshot Highlights from Task Execution: AR-VLA is also illustrated through zero-shot task execution in the real world.The section includes real-world execution screenshots.
  • E. Screenshot Highlights from Task Execution: AR-Actor specialist task execution is shown as an additional execution example.The section includes screenshots of specialist behavior.
  • F. Discussion and Limitations: The discussion identifies several challenges and future research directions despite AR-VLA’s improvements in temporal consistency and history awareness.These challenges motivate the subsequent limitations and extension proposals.
  • F. Discussion and Limitations: Autoregressive policies are sensitive to OOD trajectories because their past predictions enter the kinematic history through the Key-Value cache.A slightly OOD action can therefore trigger compounding errors rather than being reset at the next frame.
  • F. Discussion and Limitations: Action-generation gradients may fail to improve VLM semantics and can degrade pretrained semantic priors, motivating knowledge insulation.The discussion relates this phenomenon to observations in flow-matching VLAs.
  • F. Discussion and Limitations: A fully integrated autoregressive VLM could couple semantic intent more tightly with physical execution, whereas the modular Action Expert supports higher control frequencies and independent pretraining.The paper contrasts this possibility with its separate Action Expert design.
  • F. Discussion and Limitations: AR-VLA’s streaming principles could extend to VLMs by continuously refreshing vision tokens in the Key-Value cache instead of processing fixed screenshot batches.This proposed Streaming VLM aims to improve temporal resolution and real-time adaptability.
Loading 2603.10126v2…