Research project · 2026

SA-VLA: Spatially-Aware Flow-Matching for Vision-Language-Action Reinforcement Learning

Xu Pan^1,2Zhenglin Wan³Xingrui Yu^2,*Xianwei Zheng¹Youkai Ke¹Ming Sun⁴Rui Wang⁴Ziwei Wang⁵Ivor Tsang²

State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing (LIESMARS), Wuhan University
Centre for Frontier AI Research (CFAR), Institute of High Performance Computing (IHPC), Agency for Science, Technology and Research (A*STAR)
Department of Computer Science, National University of Singapore
Institute of Information Engineering, Chinese Academy of Sciences
School of Electrical and Electronic Engineering, Nanyang Technological University

^* Corresponding author

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Workshop on 3D-LLM/VLA

Abstract

RL fine-tuning can improve the in-distribution performance of flow-matching Vision-Language-Action policies while eroding their spatial generalization under viewpoint and layout shifts. We trace this failure to representation drift, sparse step-level supervision, and spatially unstructured exploration. SA-VLA restores spatial inductive bias by aligning implicit spatial representations, geometry-consistent dense rewards, and SCAN—spatially-conditioned annealed noise—in a shared latent space. Across cluttered manipulation benchmarks, it improves robustness under spatial shifts while maintaining in-distribution performance.

Spatial inductive bias collapse and preservation in SA-VLA — **Fig. 1.** Naive RL collapses spatial grounding; SA-VLA preserves it under identical perturbations.

Method

Overview of the SA-VLA framework — **Fig. 2.Overview of SA-VLA.** Visual and spatial tokens are fused into geometry-aware embeddings, then optimized through step-level dense rewards and spatially-conditioned exploration (SCAN) to preserve spatial inductive bias during RL adaptation.

Spatial token fusion

Implicit multi-view spatial tokens are projected into the visual embedding space and fused before the frozen VLM. The asymmetric attention path and gated residual preserve pretrained semantics while adding layout and view cues.

SCAN exploration

SCAN predicts action-noise scale from the geometry-aware latent and maintains an annealed lower bound. It explores uncertain, contact-sensitive states without letting stochasticity collapse during training.

Dense spatial reward

A phase-conditioned reward decomposes manipulation into Approach, Relation–Actuation, and Stabilize. It provides geometric credit assignment from relative distances, contact state, and task progress.

Instruction example	Family	Goal specification
Pick the akita black bowl between the plate and ramekin and place it on the plate.	place	`place[On](akita_black_bowl_1 → plate_1)`
Open the middle layer of the drawer.	interact	`interact[Open](wooden_cabinet_1_middle_region)`
Turn on the stove.	interact	`interact[TurnOn](flat_stove_1)`
Turn on the stove and put the moka pot on it.	multi-goal	`interact[TurnOn](flat_stove_1); place[On](moka_pot_1 → flat_stove_1_cook_region)`

Table 1. Examples of unified goal modeling used by the dense reward module.

Evaluation

We evaluate on LIBERO^[3] and LIBERO-PLUS^[4], with camera-view (CV) and robot-initial-state (RIS) shifts that respectively alter observation geometry and action feasibility. Few-shot adaptation uses three demonstrations per task-shift pair, yielding 60 evaluation configurations. All models are initialized from a $\pi_{0.5}$ flow-matching policy and optimized with PPO; the VLM backbone remains frozen.

+2.25 ppzero-shot total gain from spatial token fusion before RL

90.14%overall robustness for SA-VLA across perturbations

83.75%few-shot success in the full-model ablation setting

Method	CV	RIS	Total
$\pi_{0.5}$ baseline	78.47	83.77	81.00
SA-VLA w/ spatial fusion	82.30	84.29	83.25

Table 2. Zero-shot success rate (%) on the LIBERO-PLUS^[4] spatial-perturbation subset, before RL adaptation.

Method	SCAN	DR	Fusion	SR (%)
$\pi_{0.5}$ baseline	–	–	–	81.00
w/o Fusion	–	✓	–	81.75
w/o DR	–	–	✓	77.50
w/o SCAN	–	✓	✓	83.00
SA-VLA	✓	✓	✓	83.75

Table 3. Leave-one-out ablation on the LIBERO-PLUS^[4] few-shot spatial-perturbation setting.

Results

We separate three questions: whether dense feedback stabilizes optimization, whether structured noise improves coverage under sparse data, and whether the resulting policy transfers across heterogeneous perturbations. The sequence below follows that progression.

Training dynamics on the spatial-perturbation subset — **Fig. 4.** Training dynamics on the LIBERO-PLUS spatial-perturbation subset. Dense rewards improve convergence stability and generalization under both few-shot (solid) and zero-shot (dashed) evaluation compared with sparse supervision.

**Fig. 5.** Training dynamics under limited spatial coverage. Dense rewards stabilize optimization, while SCAN further improves performance with reduced variance. Shaded regions denote standard deviation over three seeds.

**Fig. 6.** Few-shot evaluation on LIBERO-PLUS comparing SDE-based and learned exploration noise under sparse and dense rewards. Learned, policy-dependent noise achieves higher success rates with lower variance.

Robustness improvement over pi 0.5 — **Fig. 7.** Robustness improvement over $\pi_{0.5}$ in percentage points. CV and RIS are few-shot spatial shifts; the other conditions are zero-shot. SA-VLA provides consistent gains with minimal regressions.^[1]

Methods	Total	BT	CV^†	LI	LC	OL	RIS^†	SN
$\pi_{0.5}$	88.96	100.0	82.16	88.63	98.54	100.0	83.74	87.93
ReinFlow	87.92 ± 0.81	97.53 ± 0.53	84.32 ± 0.94	84.20 ± 1.45	97.32 ± 1.69	100.0 ± 0.00	79.15 ± 1.58	92.72 ± 2.32
SA-VLA	90.14 ± 0.35	100.0 ± 0.00	87.21 ± 0.31	86.10 ± 1.52	98.54 ± 1.26	100.0 ± 0.00	83.09 ± 0.57	93.30 ± 1.85

Table 4. Robustness under diverse perturbations. We evaluate $\pi_{0.5}$^[1], ReinFlow^[2], and SA-VLA. Results report success rate in % (mean $\pm$ std). CV and RIS are few-shot spatial shifts; others are zero-shot. Total is averaged across all conditions. ^† Few-shot spatial shift.

**Fig. 8.** Phase-wise dense reward visualization during execution. Reward rises with geometric progress toward the object or target region and falls when actions depart from the intended spatial trajectory.

In motion

Representative rollouts under spatial perturbations. Eight keyframes are uniformly sampled from each successful episode; despite camera-view and robot-initial-state variation, the policy maintains coherent task progression and completes the intended manipulation.

Citation

@inproceedings{pan2026savla,
  title={SA-VLA: Spatially-Aware Flow-Matching for Vision-Language-Action Reinforcement Learning},
  author={Xu Pan and Zhenglin Wan and Xingrui Yu and Xianwei Zheng and Youkai Ke and Ming Sun and Rui Wang and Ziwei Wang and Ivor Tsang},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  year={2026},
  url={https://arxiv.org/abs/2602.00743}
}

References

Physical Intelligence, K. Black, N. Brown, et al. “$\pi_{0.5}$: A Vision-Language-Action Model with Open-World Generalization.” arXiv:2504.16054, 2025.arXiv
T. Zhang, C. Yu, S. Su, and Y. Wang. “ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning.” NeurIPS, 2025.arXiv
B. Liu, Y. Zhu, C. Gao, et al. “LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning.” NeurIPS, 2023.arXiv
S. Fei, S. Wang, J. Shi, et al. “LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models.” arXiv:2510.13626, 2025.arXiv