PaperlayerView exact revision on arXiv

Source-linked AI summary

CronusVLA: Towards Efficient and Robust Manipulation via Multi-Frame Vision-Language-Action Modeling

Hao Li, Shuai Yang, Yilun Chen, Xinyi Chen, Xiaoda Yang, Yang Tian, Hanqing Wang, Tai Wang, Dahua Lin, Feng Zhao, Jiangmiao Pang

arXiv:2506.19816v2cs.ROcs.CV

TL;DR

Single-frame VLA models underuse temporal histories, while directly feeding multiple frames into VLM backbones adds computational overhead and inference latency. CronusVLA pairs single-frame action-token pretraining with multi-frame feature post-training and chunking; it reports strong results across simulated and real-world settings, including 70.9% average success on SimplerEnv.

  • Problem

    Most VLA models use only a current observation, and direct multi-frame inputs incur high computation and inference cost.

    Source: §1 · pp. 1–2

  • Method

    CronusVLA pretrains a single-frame VLA with discrete action tokens, then post-trains feature chunks that aggregate historical frames and reuse prior features through a FIFO queue.

    Source: §3.1 · p. 3, §3.2 · p. 3

  • Results

    Across simulated and real-world settings, CronusVLA shows effectiveness and robustness, with 70.9% average success on SimplerEnv.

    Source: §1 · p. 2, §5 · p. 7

  • Takeaways & Limitations

    Experiments across simulated and real-world settings validate the effectiveness and robustness of the approach.

    Source: §5 · p. 7

    When observations and instructions remain largely unchanged, LLMs process redundant tokens; reducing inter-frame redundancy with cache-like mechanisms remains a future direction.

    Source: I Limitation and Future Works · p. 34

Abstract

from arXiv · show

Recent vision-language-action (VLA) models built on pretrained vision-language models (VLMs) have demonstrated strong performance in robotic manipulation. However, these models remain constrained by the single-frame image paradigm and fail to fully leverage the temporal information offered by multi-frame histories, as directly feeding multiple frames into VLM backbones incurs substantial computational overhead and inference latency. We propose CronusVLA, a unified framework that extends single-frame VLA models to the multi-frame paradigm. CronusVLA follows a two-stage process: (1) Single-frame pretraining on large-scale embodied datasets with autoregressive prediction of action tokens, establishing an effective embodied vision-language foundation; (2) Multi-frame post-training, which adapts the prediction of the vision-language backbone from discrete tokens to learnable features, and aggregates historical information via feature chunking. CronusVLA effectively addresses the existing challenges of multi-frame modeling while enhancing performance and observational robustness. To evaluate the robustness under temporal and spatial disturbances, we introduce SimplerEnv-OR, a novel benchmark featuring 24 types of observational disturbances and 120 severity levels. Experiments across three embodiments in simulated and real-world environments demonstrate that CronusVLA achieves leading performance and superior robustness, with a 70.9% success rate on SimplerEnv, a 26.8% improvement over OpenVLA on LIBERO, and the highest robustness score on SimplerEnv-OR. These results highlight the potential of efficient multi-frame adaptation in VLA models for more powerful and robust real-world deployment.

1 Introduction

Single-frame VLA models leave temporal history underused, while directly processing multiple images raises computational and inference costs. CronusVLA adapts single-frame pretraining through multi-frame post-training and is reported to improve benchmark performance and observational robustness.

1 Introduction

  • Prior low-level policies use histories for motion cues and state disambiguation, and consistent past observations can support action inference when the current input is corrupted.These are the two advantages of multi-frame information identified in the introduction.Source: §1 · p. 1
  • Existing VLA models typically train on one current observation, while direct multi-image inputs increase backbone computation and slow inference.The passage attributes the computation increase to self-attention scaling quadratically with input-token count and notes redundant visual tokens degrade inference speed.Source: §1 · p. 1
  • CronusVLA first trains a single-frame VLA, then post-trains it with learnable features and feature chunking to aggregate historical motion cues efficiently.The method also introduces a feature modulator and multi-frame regularization to mitigate temporal imbalance and enhance convergence.Source: §1 · p. 2
  • SimplerEnv-OR evaluates observational robustness under temporal and spatial disturbances, covering 24 disturbance types, over 120 severity levels, and 2,300 trials.The passage contrasts this evaluation with existing benchmarks' focus on tasks, objects, and scenes.Source: §1 · p. 2
  • CronusVLA averages 70.9% success on SimplerEnv, improves overall LIBERO performance by 26.8% over OpenVLA, and surpasses prior approaches on SimplerEnv-OR robustness score and success rate.The introduction also reports strong robustness in real-world experiments, without a complete success-rate figure in the supplied passage.Source: §1 · p. 2

2 Related Works

Prior VLA approaches either generate discretized actions through pretrained vision-language models or add action heads and train continuous action prediction from scratch. CronusVLA instead transfers a single-frame pretrained VLA into multi-frame modeling during post-training.

2 Related Works

  • VLA methods either adapt pretrained VLMs to generate discretized action tokens or add action heads and train continuous action prediction from scratch.CronusVLA's stated focus is transferring a single-frame, discretely pretrained VLA to multi-frame modeling with continuous action prediction.Source: §2 · p. 2
  • Low-level policies model multiple images or use video pretraining, but applying such strategies directly to large-scale VLAs introduces considerable computational overhead.Other approaches include direct multi-image input, visual prompting, and memory-based LSTMs.Source: §2 · p. 2
  • CronusVLA's framework aggregates learnable features from preceding frames in a cross-frame decoder and uses a queue to reuse feature chunks for fast inference.The framework overview also depicts single-frame pretraining followed by multi-frame post-training.Source: Figure 2 · p. 3

3 Methodology

CronusVLA establishes a single-frame VLA foundation, then adapts it to multi-frame inputs by chunking learnable features and decoding actions across frames. The section also introduces SimplerEnv-OR to evaluate robustness to spatial and temporal observational disturbances.

3.1 Single-frame Pretraining

  • Single-frame pretraining predicts discretized action tokens from one image and an instruction, transferring visual perception to embodied scenes while preserving single-frame perception at lower training cost.The extended action tokenizer maps continuous actions into 256 bins.Source: §3.1 · p. 3

3.2 Multi-frame Post-traing

  • Historical learnable features are chunked across frames and reused through a FIFO queue at inference, while a cross-frame decoder predicts action chunks from the resulting multi-frame information.The backbone processes B × M single-frame inputs per training iteration; the decoder uses diffusion loss and modulates current features to balance current and past features.Source: §3.2 · p. 3
  • Past learnable features serve as decoder-only auxiliary inputs whose gradients do not update the vision-language backbone, regularizing multi-frame training while retaining its single-frame training logic.The loss is a multi-step denoising diffusion loss.Source: §3.2 · p. 4

3.3 SimplerEnv-OR Benchmark

  • SimplerEnv-OR evaluates Bridge-v2-trained models on WR-VM tasks under spatial and temporal camera disturbances, including global, local, discrete, constant, cyclic, and sparse settings.Its R-Score measures success-rate retention relative to original WR-VM tasks; each setting includes 200 to 400 trials.Source: §3.3 · p. 4

4 Experiment

CronusVLA performs strongly across simulated and real-world manipulation, including long-horizon and disturbance-heavy tasks. Ablations examine post-training strategies, decoder designs, and the number of input frames.

4.1 Results in Simulation

  • On SimplerEnv’s Google Robot setting, CronusVLA-7B achieves the highest average success rate, scoring 78.6 in Visual Matching (VM) and 73.8 in Variant Aggregation (VA).Source: §4.1 · pp. 4–5
  • 97.0% average success on LIBERO is reported for CronusVLA with an extra wrist-view input, including 94.0% on Long, +40.3% over OpenVLA.The model performs well across almost all four task suites.Source: §4.1 · p. 5, Table 2 · p. 5
  • On SimplerEnv-OR, CronusVLA shows strong resistance to constant temporal disturbances and reaches a 96.2 R-Score under sparse 1:3 disturbances.It also consistently outperforms other models under the reported global and local spatial disturbances.Source: Table 3 · p. 5, §4.1 · p. 6

4.2 Real-world Experimental Results

  • In Franka real-world tasks, CronusVLA outperforms other models across almost all tasks, including long-horizon manipulation and generalization or disturbance tests.It withstands camera occlusions and performs best in distracted situations.Source: §4.2 · p. 6, Figure 4 · p. 6

4.3 Ablations Study

  • With total 7 frames, directly feeding multiple images into the vision-language backbone and discrete-action post-training gains only +4.5% but reduces inference speed 40.3%; decoder-only continuous feature prediction reaches 8.73 Hz versus 3.09 Hz.Training both the decoder and vision-language backbone under multi-frame post-training also improves overall performance.Source: §4.3 · p. 7
  • Cross-attention outperforms self-attention and scales linearly with frame count, while omitting the modulator degrades performance.An MLP decoder has a 20% success-rate drop; a SiT decoder has a lower average score than DiT.Source: §4.3 · p. 7, Table 5 · p. 7
  • Success rates initially rise and then fall as input frames increase; CronusVLA 7B performs best with 7 frames and 0.5B with 4.CronusVLA maintains high inference speed across frame counts and is largely unaffected by longer horizons, unlike baseline models.Source: §4.3 · p. 7, Figure 5 · p. 7

5 Conclusion

CronusVLA extends single-frame VLA models to multi-frame manipulation through two-stage training. Feature chunking and multi-frame regularization accompany fast inference, strong performance, and enhanced robustness; SimplerEnv-OR evaluates robustness under diverse disturbances.

5 Conclusion

  • CronusVLA extends single-frame VLA models to the multi-frame paradigm through a two-stage training process, with feature chunking and multi-frame regularization.The paper reports fast inference, strong performance, and enhanced robustness, and introduces SimplerEnv-OR to evaluate observational robustness under diverse disturbances.Source: §5 · p. 7

Supplementary Materials

SimplerEnv-OR evaluates observational robustness through spatial and temporal camera disturbances on WidowX pick-and-place tasks. Its scope excludes common training augmentations such as random resized crops and color adjustments.

A.1 Benchmark Settings

  • The benchmark varies spatial disturbance category and disturbed-to-clean ratios, including Constant, Cyclic, and Sparse temporal patterns.The supported ratios include 1:0, 1:1, and 1:3 or 1:5 for the respective temporal patterns.Source: Figure 6 · p. 12, Table 6 · p. 12
  • SimplerEnv-OR tests four WidowX pick-and-place tasks under disturbances without additional fine-tuning, reporting average success rate and R-Score.The evaluation uses models trained on BridgeV2, and identical random seeds ensure baselines encounter the same disturbances.Source: §A.1 · p. 12
  • Common training augmentations such as random resized cropping and brightness, contrast, saturation, and hue adjustments are outside the benchmark’s scope.The designed tasks cannot be directly solved through those simple augmentation techniques.Source: §A.1 · pp. 12–13

A.2 Spatial Evaluations on SimplerEnv-OR

On SimplerEnv-OR, CronusVLA performs strongly across spatial and temporal observational disturbances, while the benchmark’s fixed-frequency temporal perturbations leave targeted keyframe testing as a future improvement. The discussion also emphasizes the benchmark’s transferability, reproducibility, and incomplete disturbance coverage.

A.2 Spatial Evaluations on SimplerEnv-OR

  • Global, Local, and Discrete disturbances respectively affect the whole image, confined regions, or randomly distributed non-contiguous pixels.The spatial evaluation injects these disturbances into clean observations to simulate camera interference.Source: §A.2 · p. 13
  • CronusVLA achieves the highest success rate under nearly all spatial disturbance types and superior R-Score under occlusions, jittering, and blurring.Figure 7 plots average success rates and corresponding R-Scores across Global, Local, and Discrete categories; detailed results show advantages in Global and Local settings and strong robustness in Discrete settings.Source: §A.2 · pp. 13, 15, Figure 7 · p. 13

A.3 Temporal Evaluations on SimplerEnv-OR

  • Under Full Occlusion, success rates generally improve as disturbance ratios decrease, while most models approach zero success under Cyclic disturbance and CronusVLA uses past observations to infer decisions.Table 10 reports success rates across disturbance ratios alongside the original SimplerEnv results.Source: §A.3 · p. 15, Table 10 · p. 15
  • CronusVLA maintains higher success under Constant disturbances and remains largely unaffected under Sparse disturbances in the reported qualitative examples.The examples contrast its behavior with other models under Jittering and Impulse Noise.Source: §A.3 · p. 16, Figure 9 · p. 16
  • Because SimplerEnv-OR uses fixed-frequency disturbances, targeted perturbations on keyframes are a proposed improvement for more accurate robustness evaluation.Higher disturbance ratios do not always reduce success, partly because critical moments and trial variability affect outcomes.Source: §A.3 · p. 16, Figure 10 · p. 17

A.4 More Discussions

  • Applying disturbances directly to rendered image arrays decouples the benchmark from simulator modifications and supports use with standard benchmarks.Disturbance types and intensities are customizable, and deterministic trial-id seeds give models identical disturbances during evaluation.Source: §A.4 · p. 17
  • The authors identify incomplete disturbance coverage and plan to expand baselines, simulation environments, and evaluation protocols.They also encourage robustness evaluation for VLA deployment and future applications.Source: §A.4 · p. 17

B.1 Simulation Experiments

In simulation, CronusVLA performs strongly on the long-horizon Put in Drawer task, completing steps that comparison models fail to finish reliably. In a real-world button-pressing task, its multi-frame context helps distinguish ambiguous states and execute the instructed sequence.

B.1 Simulation Experiments

  • CronusVLA reaches 64.8% and 65.1% on Put in Drawer in Visual Matching and Variant Aggregation, respectively, versus baseline success rates of 0%–30%.In the illustrated trials, OpenVLA stalls while opening the drawer, and RoboVLMs repeatedly collides after initially opening it; CronusVLA completes all steps.Source: §B.1 · p. 19, Figure 11 · p. 19

B.2 Real-world Experiments

  • In the real-world button-sequencing task, CronusVLA executes the red-yellow-green order, while OpenVLA repeats presses because similar observations recur before and after presses.The authors attribute CronusVLA’s performance to task decomposition, instruction following, and temporal context for resolving ambiguous states.Source: §B.2 · p. 20, Figure 12 · p. 20

C.1 More Explorations on Multi-frame Training

Single-frame pretraining and multi-frame regularization support more efficient multi-frame training, while the experiments examine feature choice, post-training data, and results across SimplerEnv and LIBERO. The reported evaluations show advantages across robots and comparable LIBERO performance to OpenVLA-OFT with fewer training steps.

C.1 More Explorations on Multi-frame Training

  • CronusVLA scores 70.9 on SimplerEnv after 50k steps, while scratch-trained models score 10.4 and 18.1 under the two reported settings.The scratch models fail to converge within the limited budget; without multi-frame regularization, scratch settings show lower average success and a tendency toward slower convergence.Source: §C.1 · p. 21, Figure 13 · p. 21

C.2 Ablation on Types of Learnable Feature

  • The final learnable-feature strategy is adopted because multi-layer features raise complexity and extra visual tokens degrade performance.Multi-layer final improves performance but reduces usability; the chosen strategy uses a last-layer feature placed after all visual and language tokens.Source: §C.2 · p. 21, Table 12 · p. 21

C.3 Ablation on Post-training Data

  • Joint post-training on Bridge-v2 and Fractal consistently improves performance across the WidowX and Google robots compared with single-dataset training.The authors attribute this pattern to complementary datasets and more robust, generalized representations.Source: §C.3 · p. 22, Table 13 · p. 22

C.4 Detailed Main Result in SimplerEnv

  • CronusVLA 0.5B and 7B outperform prior baselines in most average SimplerEnv task results across Google and WidowX settings.The reported detailed results cover Google Robot success rates and WidowX Robot success rates.Source: §C.4 · p. 22, Table 14 · p. 22, Table 15 · p. 23

C.5 More Results in LIBERO Benchmark

  • CronusVLA achieves performance comparable to OpenVLA-OFT on LIBERO without state inputs or additional optimization, using 30K rather than 150K training steps.The results cover four task suites; Table 16 reports success rates and standard errors averaged over three seeds and 500 trials.Source: §C.5 · p. 23, Table 16 · p. 23

D Training Details

Training proceeds from single-frame pretraining through multi-frame post-training and task-specific finetuning. The model combines a VLA module, decoder, and cached feature chunks for multi-frame inference.

D.1 Pretraining Details

  • Single-frame pretraining uses 27 Open X-Embodiment datasets for the 7B model and cross-entropy loss on predicted action tokens.The same procedure is applied to the 0.5B model using its specified VLM backbone.Source: §D.1 · p. 24

D.2 Post-training Details

  • Post-training uses Fractal and Bridge-v2 data, end-to-end noise prediction, and a 100-step diffusion schedule without autoregressive text-token prediction.Data processing includes RLDS preprocessing and image augmentations; the data loader samples across datasets, trajectories, and time steps.Source: §D.2 · p. 24
  • CronusVLA-7B post-training uses 6-frame histories, while the 0.5B model uses 3-frame histories and a 16-action chunk with adaptive ensembling.The 7B model trains for approximately 50k steps on 64 A100 GPUs; the 0.5B model trains for about 40k steps on 32 A100 GPUs.Source: §D.2 · p. 24

D.3 Finetuning Details

  • Task-specific finetuning adapts post-trained weights to expert demonstrations, using 3-frame histories and action chunks of length 8.CronusVLA-7B is finetuned for 6K–30K steps.Source: §D.3 · p. 24

E.1 Model Architecture

  • The architecture pairs a vision-language module with a 12-layer decoder that uses self-attention and cross-attention to connect noised actions with learnable features.The vision encoder combines SigLIP and DINOv2; actions and timesteps are embedded in 768 dimensions.Source: §E.1 · p. 25, Table 17 · p. 25

E.2 Feature Chunking

  • Feature chunking reuses cached learnable features in a FIFO queue, padding early histories with repeated first-frame features to keep feature length consistent.The same padding is introduced through stochastic sampling during training.Source: §E.2 · p. 25, Figure 15 · p. 25

E.3 Efficiency Analysis

CronusVLA maintains 8.7 Hz inference as its history grows from 2 to 7 frames, and the real-world evaluation uses a Franka robot platform with deployment constraints.

E.3 Efficiency Analysis

  • 8.7 Hz inference speed remains constant as CronusVLA-7B history length increases from 2 to 7 frames.The method also supports a compact 0.5B model and a larger 7B model; post-training requires approximately 50K steps.Source: §E.3 · p. 26, Table 18 · p. 26

F.1 Experimental Setup

F.2 Tasks Settings

The real-world evaluation covers simple pick-and-place, multi-step manipulation, and generalization and robustness conditions. It also documents representative failure modes in real-world and simulated tasks.

F.2 Tasks Settings

  • Simple pick-and-place tasks test object localization, grasping, placement precision, spatial generalization, and instruction following.The tasks include object picking, carrot placement, cube stacking, and cup stacking, with success rates reported in Table 19.Source: §F.2 · p. 27, Table 19 · p. 28
  • Long-horizon tasks assess sequential execution through multi-object placement, drawer manipulation, drawer-and-carrot placement, and ordered button pressing.The drawer tasks vary initial opening states; the button task tests disambiguation across visually similar stages, with results reported in Table 20.Source: §F.2 · p. 28, Table 20 · p. 28
  • Generalization and robustness tests vary carrot materials, introduce human interference and occlusion, and use unseen objects, backgrounds, distractors, and lighting.Policies are evaluated on carrot placement after training data from all seven simple pick-and-place tasks; results are reported in Table 21.Source: §F.2 · pp. 28–29, Table 21 · p. 29

G Failure Cases

  • Observed failures include unnecessary grasp adjustments, inaccurate depth estimation for distant drawers, occlusion-related collisions, and unrealistic simulated contact outcomes.The authors identify single-view perception, obstacle avoidance under occlusion, and a real-to-sim gap as areas for improvement.Source: G Failure Cases · p. 33, Figure 20 · p. 33

H Complexity Analysis

The complexity analysis contrasts quadratic attention growth for directly concatenated frames with linear growth for feature-level modeling. The paper also describes scope boundaries, broader impact, and supporting data and benchmark setups.

H Complexity Analysis

  • Directly concatenating M past frames raises VLM self-attention complexity from O((P + I)^2) to O(((M + 1) · P + I)^2), approximately O(M^2) when P ≫ I.The proposed approach processes frames independently through the VLM and models their relations at the feature level.Source: H Complexity Analysis · p. 34
  • Feature-level modeling has complexity O(M · TVLM) + O(M · Tdecoder), approximately O(M), because decoder inference time is much smaller than VLM inference time.The authors state that this linear scaling has limited impact on inference speed in practice.Source: H Complexity Analysis · p. 34

I Limitation and Future Works

  • CronusVLA uses single-view images and language instructions, without proprioceptive states or multi-view observations, and does not explicitly use foundational language abilities during inference.Future directions include cache-like use of inter-frame differences, language-driven reasoning, and multi-view and state-conditioned temporal modeling.Source: I Limitation and Future Works · p. 34

J Broader Impact

  • The paper identifies assistive robotics and industrial settings as potential beneficiaries of improved temporal understanding for long-horizon manipulation.It frames the potential benefit as better interpretation and execution of complex, time-extended instructions.Source: J Broader Impact · p. 34

K More Backgrounds

  • The 7B vision-language backbone combines LLaMA2 with DINOv2 and SigLIP, and its Prismatic alignment training uses roughly 558K image-caption pairs.Prismatic is further instruction-tuned on 665K multimodal examples.Source: K More Backgrounds · p. 35
  • VLA pretraining uses 27 Open X-Embodiment datasets, including Fractal and Bridge-v2 demonstrations stored in RLDS format.The cited dataset descriptions note possible imperfect instructions in some Kuka subsets and potential out-of-distribution issues in Droid and Language Table.Source: K More Backgrounds · p. 35
  • Continuous action dimensions are divided into 256 uniform bins and mapped to discrete tokens, with cross-entropy used for next-token prediction.The 256 least-frequently used Llama tokenizer tokens are repurposed for action tokens.Source: K More Backgrounds · p. 35
  • SimplerEnv evaluates WidowX and Google Robot tasks, while LIBERO comprises four suites covering spatial, object, goal-conditioned, and long-horizon manipulation.Each LIBERO suite contains 10 tasks and 50 human-teleoperated demonstrations.Source: K More Backgrounds · p. 35, Figure 21 · p. 36, Figure 22 · p. 37
Loading 2506.19816v2…