PaperlayerView exact revision on arXiv

Source-linked AI summary

FLOWER: Democratizing Generalist Robot Policies with Efficient Vision-Language-Action Flow Policies

Moritz Reuss, Hongyi Zhou, Marcel Rühle, Ömer Erdinç Yağmurlu, Fabian Otto, Rudolf Lioutikov

arXiv:2509.04996v1cs.RO

TL;DR

Efficient generalist VLAs must balance semantic VLM capacity, expressive action generation, and practical computational demands. FLOWER addresses this with intermediate fusion and action-space Global-AdaLN in a compact flow-based policy, achieving competitive performance across broad evaluations while retaining documented scope limitations.

  • Problem

    Existing VLA policies require substantial model capacity and computation, while allocating capacity between semantic VLM processing and expressive diffusion-based action modeling remains difficult.

  • Method

    FLOWER prunes 30–50% of VLM layers for intermediate-modality fusion and uses action-space Global-AdaLN conditioning in a 950M-parameter flow-based VLA.

  • Results

    FLOWER matches current state-of-the-art VLAs across 190 tasks in 10 benchmarks, while achieving 61% average real-world success versus OpenVLA’s 31%.

  • Takeaways & Limitations

    FLOWER demonstrates that a compact VLA can deliver competitive generalist manipulation performance with 950M parameters and 200 GPU hours of pretraining.

  • Takeaways & Limitations

    Validation primarily covers three manipulation action spaces; generalization to mobile navigation or humanoid locomotion remains unexplored.

Abstract

from arXiv · show

Developing efficient Vision-Language-Action (VLA) policies is crucial for practical robotics deployment, yet current approaches face prohibitive computational costs and resource requirements. Existing diffusion-based VLA policies require multi-billion-parameter models and massive datasets to achieve strong performance. We tackle this efficiency challenge with two contributions: intermediate-modality fusion, which reallocates capacity to the diffusion head by pruning up to $50\%$ of LLM layers, and action-specific Global-AdaLN conditioning, which cuts parameters by $20\%$ through modular adaptation. We integrate these advances into a novel 950 M-parameter VLA called FLOWER. Pretrained in just 200 H100 GPU hours, FLOWER delivers competitive performance with bigger VLAs across $190$ tasks spanning ten simulation and real-world benchmarks and demonstrates robustness across diverse robotic embodiments. In addition, FLOWER achieves a new SoTA of 4.53 on the CALVIN ABC benchmark. Demos, code and pretrained weights are available at https://intuitive-robots.github.io/flower_vla/.

1 Introduction

FLOWER addresses the efficiency–capacity trade-off in generalist VLA policies by reallocating capacity from pruned VLM layers to flow-based action modeling. The resulting 950M-parameter policy targets competitive performance across diverse tasks and settings with substantially lower training cost.

  • Motivation: Existing diffusion-based VLAs face a capacity trade-off: large VLM backbones limit diffusion-head expressivity, while smaller VLMs lose semantic features and full-VLM denoising slows training and inference.These constraints complicate efficient deployment across diverse robotic tasks and embodiments.
  • Contributions: 30–50% of pretrained VLM layers are pruned, and intermediate embeddings condition a Flow Transformer to retain semantic grounding while reclaiming parameters.Intermediate fusion reallocates capacity without discarding all VLM representations.
  • Contributions: 20% fewer diffusion-head parameters are achieved with action-space Global-AdaLN while preserving accuracy and expressivity.The mechanism uses action-specific normalization within the diffusion transformer.
  • FLOWER: 950M parameters and 99% lower pretraining cost support competitive performance across 190 tasks in 10 simulation and real-world benchmarks.The paper presents FLOWER as a compact VLA policy designed for heterogeneous robotics data.

2 Related Work

Related work develops generalist robotic policies through larger datasets, pretrained vision-language models, diffusion heads, and alternative fusion strategies. FLOWER builds on these directions while focusing on intermediate fusion for efficient VLA design.

  • Generalist policy learning: Large-scale imitation-learning datasets, including OXE’s 1.4M trajectories across more than 20 embodiments, support research on generalist policies.Prior diffusion-policy methods also train directly on heterogeneous robotic data without pretrained VLMs.
  • Vision-language-action policies: OpenVLA fine-tunes a 7.7B-parameter VLM for discrete action prediction, but its size makes real-robot deployment challenging.Other VLA approaches use discrete actions, asynchronous diffusion, or latent commands conditioning diffusion policies.
  • Fusion strategies: Fusion strategies range from early or late fusion to FLOWER’s intermediate fusion, which injects mid-level VLM tokens into the Flow Transformer.Intermediate fusion is presented as an alternative to combining streams only at the input or after separate processing.

3 Method

FLOWER combines intermediate VLM–Flow Transformer fusion, action-space Global-AdaLN conditioning, and rectified flow to model heterogeneous robotic actions efficiently. Its compact architecture uses a pruned VLM, cross-attention, metadata conditioning, and a diverse multi-dataset pretraining mixture.

  • Method formulation: FLOWER learns actions conditioned on robot state, language goals, and embodiment metadata across heterogeneous action, observation, and task spaces.The policy is trained as a unified generalist model over trajectories from multiple embodiments.
  • Intermediate modality fusion: Intermediate VLM features are extracted after multimodal encoding and injected into the Flow Transformer through cross-attention.The design aims to preserve semantic context while reducing computational cost and improving convergence.
  • Cross-action-space Flow Transformer: Action-Space Global-AdaLN-Zero conditions each transformer block on temporal signals and action-type embeddings for heterogeneous action spaces.A shared modulation set generates action-specific signals across layers, with lightweight LoRA adapters adding capacity.
  • Rectified flow for action generation: Rectified flow models straight-line velocity fields between Gaussian noise and action data, requiring 4 denoising steps for single-arm and 8 for high-frequency dual-arm settings.The flow model is conditioned on state, language, and embodiment during action generation.
  • Implementation: 947M parameters are organized into half of the Florence-2-L VLM, an 18-layer Flow Transformer, and action-specific modules.Pretraining uses an approximately 250k-trajectory mixture of eight public robotic datasets with broad scene and embodiment diversity.

4 Evaluation

FLOWER is evaluated through ablations, simulation benchmarks, and real-world manipulation and generalization experiments. Across these settings, it matches or surpasses strong baselines while demonstrating efficient inference and robust transfer.

  • Critical design decisions: Intermediate fusion achieves 93.4% success on LIBERO-Long, outperforming early fusion at 33.4% and late fusion at 73%.The comparison is reported with the Florence-VLM variant.
  • Critical design decisions: 20% parameter reduction from Global-AdaLN does not reduce performance on CALVIN ABC.The method is compared against default AdaLN in the Florence-based FLOWER model.
  • Critical design decisions: A small 384-dimensional, six-layer action head reaches an average performance of 2.60, confirming the importance of a high-capacity Flow Transformer.The result comes from an ablation of the diffusion transformer capacity.
  • Simulation experiments: FLOWER matches or surpasses current state-of-the-art approaches across reported simulation benchmarks, including wide-margin gains over OpenVLA on CALVIN and LIBERO.It also outperforms π0 on LIBERO and ALOHA, supporting adaptation across embodiments and task settings.
  • Real-world evaluation and generalization: FLOWER achieves 61% average success across 20 real-world tasks, compared with 31% for OpenVLA.Each task is evaluated five times from randomized starting positions in a Franka Panda kitchen setup.
  • Real-world evaluation and generalization: FLOWER averages 51.0% success across challenging real-world generalization scenarios, compared with 23.4% for OpenVLA.The scenarios include novel objects, flashlight-only lighting, background distractors, and new task compositions, while fine manipulation in clutter remains challenging.
  • Inference efficiency: FLOWER reaches 311Hz throughput, 8% faster than π0 and 5007% faster than OpenVLA, while using 27.6% and 12.7% of their memory footprints.The measurements were collected on an RTX 4090 GPU.

5 Conclusion

The conclusion presents FLOWER as a compact flow-based VLA that combines intermediate fusion and Global-AdaLN to reduce computational requirements. It reports competitive performance across broad task and benchmark coverage despite a 950M-parameter model and limited pretraining budget.

  • Contributions: Intermediate fusion prunes 30–50% of VLM layers, while Global-AdaLN reduces transformer-head parameters by 20% without compromising expressivity.Together, these techniques support a compact flow-based VLA design.
  • Overall results: FLOWER matches current state-of-the-art VLAs across 190 tasks in 10 benchmarks with 950M parameters and 200 GPU hours of pretraining.The conclusion frames this as competitive performance at substantially reduced model and training costs.

6 Limitations

FLOWER’s limitations concern inference speed, embodiment coverage, zero-shot performance, deployment scale, and the predominance of simulation-based evaluation.

  • Scope and deployment: Iterative sampling is inherently slower than a single forward pass from deterministic policies.This constrains inference efficiency despite FLOWER’s other efficiency advantages.
  • Scope and deployment: Validation covers primarily three manipulation action spaces, leaving mobile navigation and humanoid locomotion unexplored.The paper identifies broader embodiment generalization as future work.
  • Scope and deployment: Zero-shot deployment on the SIMPLER Google Robot benchmark still needs improvement, possibly because its generalization benefits from larger models.The larger-model explanation is presented as a hypothesis.
  • Scope and deployment: FLOWER’s approximately 1B parameters may still challenge low-resource or real-time deployment.The model is smaller than most state-of-the-art VLA models but remains substantial.
  • Scope and deployment: Eight of the ten benchmarks are conducted in simulation, limiting the evidence for real-world generalization.This is an evaluation-distribution limitation rather than a claim that real-world transfer fails.

A.1 Pretraining Ablation Experiences

The pretraining ablations identify design choices that improve efficiency, stability, and cross-action generalization. FLOWER uses constant-length action chunks, single-image pretraining, action-specific normalization, and stabilizing transformer components.

  • Ablated pretraining choices: 20-step constant action chunks provided a practical trade-off across datasets after variable-length chunks converged slowly and performed worse in SIMPLER.The fixed length also supports finetuning to different action lengths.
  • Ablated pretraining choices: Single-image pretraining avoided the speed, memory, and convergence costs introduced by flexible multi-view image inputs.Finetuning experiments retained flexibility for multiple images.
  • Ablated pretraining choices: Mixture-of-experts flow-transformer variants were more memory-intensive, converged more slowly, and encountered NaN losses.The tested design combined shared and action-specialist MLP components.
  • Optimization stability: The final architecture reduced the apparent importance of the custom scheduler: replacing it with a constant 2 × 10^-5 learning rate caused only a small change in the finalized model.The earlier 4.44 versus 0.8 drop was largely attributed to instability in an older model variant.
  • Flow Transformer design: Action-specific dual-RMSNorm parameters capture activation statistics for heterogeneous action types more effectively than a single normalization.RMSNorm, SwiGLU MLPs, QK-value normalization, and related components were used to improve training stability.

B Detailed Experiments

The experiments evaluate FLOWER against diverse baselines on CALVIN and summarize performance across benchmark suites. FLOWER surpasses all listed CALVIN baselines with six hours of finetuning, while the tables support cross-benchmark comparison.

  • Cross-benchmark comparison: The benchmark comparison spans CALVIN, real-generalization, and other benchmark-level normalized improvements relative to each benchmark’s second-best baseline.CALVIN normalization uses average sequence length divided by five, while real-generalization averages four specified tests.
  • Experimental setup: FLOWER experiments train for up to 40k steps on four GPUs with batch size eight per GPU.The standardized protocol is intended to enable direct baseline comparison.
  • CALVIN Benchmark: FLOWER is evaluated against VLA, video-based, and diffusion-based baselines across CALVIN variants.The evaluation reports task-level success rates and average rollout length for five-instruction chains.

B.1 SIMPLER Benchmark Tasks

SIMPLER tests FLOWER on diverse real-to-simulation tasks across Google Robot and WidowX + Bridge settings. FLOWER outperforms Octo and OpenVLA on both benchmarks, while RT-1X remains strongest on several Google Robot tasks.

  • Google Robot Setting: The Google Robot setting covers picking a can, relative object positioning, drawer manipulation, and sequential drawer opening with object placement.These tasks vary object configurations, robot positions, drawer choices, and multi-step requirements.
  • WidowX + Bridge Setting: The WidowX + Bridge setting tests object transfer, object substitution, precise stacking, and placing an eggplant into a basket.The tasks vary gripper reorientation, spatial constraints, object identity, and object orientation.
  • Benchmark scope: SIMPLER evaluates approximately 3,000 rollouts across eight tasks and varied conditions in the Google Robot and Bridge settings.FLOWER is pretrained on a cross-embodiment mixture before evaluation.
  • Results: FLOWER outperforms Octo and OpenVLA on both SIMPLER benchmarks after only 200 GPU hours of heterogeneous pretraining.It achieves stronger overall performance on Bridge, while RT-1X leads several Google Robot tasks and FLOWER ranks second there among the reported policies.
  • Results: The results indicate strong performance under low computational demand and robustness across diverse embodiments and action spaces, while Google Robot gaps motivate further refinement.The Bridge gains highlight handling of diverse action spaces after heterogeneous pretraining.

B.1.1 LIBERO Benchmark.

LIBERO evaluates FLOWER on multiple manipulation suites using a Franka Panda with dual-camera observations. FLOWER achieves near-perfect success across variants and is uniquely above 90% on LIBERO-Long among the generalist policies described.

  • Benchmark setup: LIBERO-10 provides 50 demonstrations for 10 tasks, while LIBERO-90 covers 90 tasks using Franka Panda end-effector control and static and wrist cameras.The suites test spatial, goal, object, and long-horizon manipulation properties.
  • Results: FLOWER significantly outperforms the listed generalist and specialist baselines across every LIBERO variant, with success rates consistently above 93%.The comparison includes π0, π0-FAST, π0.5-ki, OpenVLA variants, CoA-VLA, Octo, MiniVLA, and Baku.
  • LIBERO-Long: On LIBERO-Long, FLOWER reaches 93.5% success and is the only policy described as exceeding 90%.Other generalist approaches achieve 50–54%, while Baku is the only specialist reported as competitive in this setting.
  • Results: The LIBERO results are reported in a table using success rate as the metric, with best results marked in bold.The table caption states that FLOWER achieves state-of-the-art results across most tested settings.

B.1.2 Aloha Benchmark.

FLOWER is evaluated on challenging Aloha bimanual manipulation tasks, including high-frequency control, against specialist policies and varied pretraining mixes. It performs strongly with few denoising steps, while standard Diffusion Policy fails on both tasks and cross-embodied pretraining underperforms the joint-only mix.

  • Evaluation setup: The Aloha evaluation uses a bimanual robot with joint-space control, a top-view camera, proprioceptive input, and 500 episodes per task.
  • Results: FLOWER achieves strong performance on both Aloha simulation tasks with eight denoising steps.The tasks are Insert the peg into the socket and Pick up the object and transfer it.
  • Results: FLOWER outperforms the specialist ACT policy by a considerable margin on the challenging Insertion task.
  • Results: FLOWER achieves comparable Transfer performance to ACT except for the variant pretrained on cross-action-space data.
  • Results: The standard Diffusion Policy cannot solve either Aloha task.
  • Pretraining comparison: The joint-only Droid pretraining mix achieves the best FLOWER results, while cross-embodied pretraining performs poorly.

B.3.1 Novel Task Compositions

The paper tests compositional generalization through long-horizon sequences that require ordered subtasks, alongside evaluations of unseen objects, lighting, and distractors. FLOWER is reported to perform robustly on novel objects and to apply rectified flow as an open-source generalist component across diverse embodiments and contexts.

  • Novel Task Compositions: Novel task compositions require agents to execute multiple subtasks in a prescribed order as one coherent long-horizon plan.
  • Novel Task Compositions: The appliance sequence combines opening and closing the microwave, oven, and ice machine in a specified order.
  • Novel Task Compositions: The stovetop-to-sink sequence tests coordinated transfers, lever pushing, object placement, and repeated workstation moves.
  • Generalization analysis: The generalization table evaluates Novel Object, Flashlight, and Background Distractors scenarios using success rates across kitchen tasks.
  • Related work and positioning: Rectified flow enables few-step action sampling, and FLOWER applies it as an open-source generalist component across diverse vision-language contexts and embodiments.
Loading 2509.04996v1…