StairVLA: Stage-Aware Hierarchical Action Generation for Vision-Language-Action Models

Shangyuan Yuan; Xinda Qi; Yujiang Pu; Wenliang Guo; Xiaobo Tan

cs.ROOct 6, 2026

StairVLA: Stage-Aware Hierarchical Action Generation for Vision-Language-Action Models

Authors: Shangyuan Yuan, Xinda Qi, Yujiang Pu, Wenliang Guo, Xiaobo Tan

Organizations: Michigan State University · Ant Group

Abstract

Vision-language-action (VLA) models increasingly rely on diffusion- or flow-matching-based action heads to generate continuous robot actions. These action heads typically process the denoising trajectory in a largely uniform manner. However, we observe that the conditioning focus naturally shifts across denoising stages: early stages combine language instructions and visual observations to establish a coarse action trajectory, whereas later stages place greater emphasis on current visual observations for action alignment. Based on this insight, we introduce StairVLA, a stage-aware hierarchical action generation framework that uses partially denoised actions as a natural interface between coarse long-horizon action generation and local refinement. A high-level VLA performs early denoising to produce a reusable long-horizon partially denoised action trajectory, while a lightweight refiner operates at a higher frequency to refine local action chunks using the latest observations. This design amortizes expensive high-level VLA computation while preserving frequent closed-loop correction. On LIBERO, our GR00T-style instantiation improves average success from 96.5% to 97.8% while reducing amortized inference latency from 115.0 ms to 44.2 ms per action chunk. More broadly, across two VLA backbones, simulation benchmarks, and real-robot tasks, StairVLA consistently reduces inference cost while maintaining strong task performance.

Figures & tables

Figure 1 : Stage dependence in generative VLA action formation. (A) Conceptual illustration of the evolving conditioning focus across denoising stages, from joint instruction–observation conditioning toward stronger visual alignment. (B) Relative attention to instruction and image tokens across denoising stages in $\pi_{0}$ on LIBERO, showing an increasing emphasis on visual observations during later refinement. (C) Task success of partially denoised actions generated by GR00T on LIBERO, showing that substantial task competence emerges before denoising is complete.

Figure 2 : Overall framework of StairVLA. The high-level VLA performs early denoising over a long-horizon trajectory and caches the partially denoised trajectory for reuse across multiple steps. For each local chunk, the refiner uses the corresponding segment and the latest observation to complete refinement. The top, middle, and bottom sequences denote the initial noise, partially denoised trajectory, and executable action chunk, respectively.

Figure 3 : Overview of the low-level refiner. (A) The refiner uses the current observation, compressed high-level VLM tokens, and partially denoised action context to refine the query chunk. (B) The refiner adopts a GR00T-style DiT with interleaved cross-attention and self-attention for conditional fusion and action modeling.

Figure 4 : Refiner training path construction. After early denoising by the high-level VLA, the resulting partially denoised action is used to construct the refinement path toward the ground-truth action, from which both refiner inputs and targets are generated.

Representative VLA Policies
Method	Spatial	Object	Goal	Long	Avg.	Latency
OpenVLA ( Kim et al., 2024 )	84.7	88.4	79.2	53.7	76.5	–
$\pi_{0}$ ( Black et al., 2024 )	96.8	98.8	95.8	85.2	94.1	–
GR00T ( Bjorck et al., 2025 )	94.4	97.6	93.0	90.6	93.9	–
$\pi_{0.5}$ ( Physical Intelligence et al., 2025 )	98.8	98.2	98.0	92.4	96.9	–
Hierarchical / Coarse-to-Fine VLA Policies

Table 1 : Comparison on the LIBERO benchmark. We report success rates (%) on the four LIBERO suites, the average success rate (%), and reported inference latency (ms/chunk) when available. Chunk sizes may vary across methods. Best and second-best results are shown in bold and underlined , respectively. ∗ Uses additional subtask-level supervision beyond standard action demonstrations and is excluded from best/second-best ranking. ‡ StreamVLA reports average wall-clock latency under its gated reasoning schedule. § CF-VLA reports action-sampling latency excluding visual-language prefix encoding and KV-cache construction.

Zero-Shot Transfer
Method	Camera	Robot	Language	Light	Background	Noise	Layout	Avg.
OpenVLA ( Kim et al., 2024 )	0.8	3.5	23.0	8.1	34.8	15.2	28.5	15.6
$\pi_{0}$ -FAST ( Pertsch et al., 2025 )	65.1	21.6	61.0	73.2	73.2	74.4	68.8	61.6
OpenVLA-OFT ( Kim et al., 2025 )	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
GR00T ( Bjorck et al., 2025 )	34.6	50.8	85.1	86.5	86.5	63.0	73.6	66.8
Libra-VLA ( Wei et al., 2026 )	68.9	48.8	92.7	97.9	93.4	86.3	77.5	79.5

Table 2: Results on LIBERO-Plus under zero-shot transfer and supervised fine-tuning.

Setting	Refiner	Action Context	Spatial	Object	Goal	Long	Avg.	$\Delta$
StarVLA-GR00T ( StarVLA Community, 2026 )	–	–	97.8	98.8	97.4	92.0	96.5	–
Ours w/o refiner (Top-only)	✗	–	96.8	89.4	96.8	86.4	92.4	-4.1
Ours w/o Action Context	✓	✗	99.2	98.6	95.6	95.4	97.2	+0.7
Ours	✓	✓	98.0	99.6	97.8	95.8	97.8	+1.3

Table 3: Ablations on LIBERO.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 11 : Denoising-stage attention analysis for GR00T and

\pi_{0}

. We record action-query cross-attention to different conditioning-token groups over 100 dense denoising stages. Both models exhibit stage-dependent attention patterns, with the changes being more pronounced in the first action-head layer.

Panel of Figure 12 — Figure 12 : Representative first-layer cross-attention maps of $\pi_{0}$ and GR00T at different denoising stages. For $\pi_{0}$ , attention to image tokens becomes stronger toward later stages, while task-instruction attention does not exhibit a clear monotonic trend. For GR00T, late-stage cross-attention does not simply shift toward raw image or task-instruction tokens, indicating that different VLA action heads may realize stage-dependent conditioning through different token pathways.

Figure 13 : First-layer self-attention allocation in GR00T across denoising stages. Attention to action keys increases toward later stages, while attention to state/future keys decreases.

Method	$M$	Chunk Size	Spatial	Object	Goal	Long	Avg.	$\Delta$ Avg.	Latency
GR00T-style backbone
StarVLA-GR00T	–	8	97.8	98.8	97.4	92.0	96.5	–	115.0
Ours (GR00T-base)	2	5	98.4	98.8	97.6	94.8	97.40	+0.90	68.7
Ours (GR00T-base)	3	5	99.2	98.8	97.8	94.0	97.45	+0.95	52.3
Ours (GR00T-base)	4	5	98.0	99.6	97.8	95.8	97.80	+1.30	44.2
Ours (GR00T-base)	5	5	98.4	98.8	97.4	93.6	97.05	+0.55	39.3

Appendix

Table 4: Complete LIBERO results across GR00T-style and

\pi

-style VLA backbones.

M

denotes the number of 5-action low-level chunks executed between two high-level policy updates. The hierarchical GR00T- and

\pi

-based models use high-level action horizons of

H=32

and

H=20

, respectively. Latency is reported per native returned action chunk. The non-hierarchical StarVLA baselines use 8-action chunks, whereas our low-level refiner outputs 5-action chunks.

Figure 14 : Effect of long-horizon high-level trajectory reuse. We report the average LIBERO success rate as the cached high-level trajectory is reused for increasingly long execution horizons. Both curves are shown from 25 executed actions onward, corresponding to the shared evaluation range of the

H=48

and

H=64

models.

Task ID	Benchmark	Task Description	Retained Demonstrations
1	Fruit25	Pick up the apple and place it into the red basket.	100
2	Fruit25	Pick up the apple and place it onto the metal tray.	100
3	Fruit25	Pick up the orange and place it onto the metal tray.	100
4	Fruit25	Pick up the orange and place it into the orange basket.	96
5	Fruit25	Pick up the lemon and place it onto the metal tray.	98
6	Fruit25	Pick up the lemon and place it into the yellow basket.	99

Appendix

Table 5: Task definitions and retained demonstration counts for the real-world datasets.

Fruit25
ID	Task	StarVLA-GR00T	StarVLA- $\pi$	Ours
1	Apple $\rightarrow$ Red basket	7/10 (70%)	6/10 (60%)	10/10 (100%)
2	Apple $\rightarrow$ Metal tray	10/10 (100%)	9/10 (90%)	9/10 (90%)
3	Orange $\rightarrow$ Metal tray	9/10 (90%)	9/10 (90%)	9/10 (90%)
4	Orange $\rightarrow$ Orange basket	5/10 (50%)	8/10 (80%)	7/10 (70%)
5	Lemon $\rightarrow$ Metal tray	10/10 (100%)	10/10 (100%)	10/10 (100%)

Appendix

Table 6: Per-task success results on the real-world benchmarks. Fruit25 tasks are evaluated with either 5 or 10 trials, while PushBlock is evaluated with 10 trials per method.

Panel of Figure 17 — Figure 17 : Examples of relatively easy PushBlock initial configurations, where the block is approximately aligned with the target region and can be pushed along a near-straight trajectory.

Method	T-SPARC $\uparrow$	R-SPARC $\uparrow$	T-Jerk $\downarrow$	R-Jerk $\downarrow$	Stop (%) $\downarrow$
StarVLA-GR00T	-5.480 [-5.817, -5.127]	-10.499 [-11.932, -9.440]	9.599 [9.408, 9.788]	10.026 [9.833, 10.216]	12.537 [10.575, 15.102]
StarVLA- $\pi$	-6.799 [-7.266, -6.233]	-12.134 [-13.370, -11.135]	9.939 [9.681, 10.212]	10.355 [10.091, 10.599]	26.066 [23.566, 28.767]
Ours	-5.151 [-5.398, -4.887]	-9.992 [-11.099, -8.959]	9.484 [9.300, 9.644]	9.909 [9.733, 10.091]	11.274 [8.974, 13.607]

Appendix

Table 7: Motion smoothness and execution continuity on Fruit25 trials. Values are reported as median [Q1, Q3]. Higher SPARC and lower jerk and stop ratio indicate smoother and more continuous execution.

Explore similar work

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies
Apr 27, 2026Fan Du, Feng Yan, Jianxiong Wu +8Robotic Control Efficient VLA Model Inference
1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology +1
Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference
Jul 14, 2026Yuzhou Wu, Yuxin Zheng, Muchun Niu +6Efficient VLA Model Inference Efficient VLA Models
1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2
VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting
Jul 7, 2025Juyi Lin, Amir Taherin, Arash Akbari +11Robotic Control Efficient VLA Model Inference
Northeastern University, Boston, USA · EmbodyX,San Mateo,USA

Apr 27, 2026, cs.CVJ/K move · Enter open · S save

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies

Fan Du, Feng Yan, Jianxiong Wu, Xinrun Xu+7

1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology+1

Robotic Control Efficient VLA Model Inference Visuomotor Policy Learning Efficient VLA Models Robot Motion Generation Vision-Language-Action Models

Open paper AskPDF arXiv

Flow-based vision-language-action (VLA) policies offer strong expressivity for action generation, but suffer from a fundamental inefficiency: multi-step inference is required to recover action structure from uninformative Gaussian noise, leading to a poor efficiency-quality trade-off under real-time constraints. We address this issue by rethinking the role of the starting point in generative action modeling. Instead of shortening the sampling trajectory, we propose CF-VLA, a coarse-to-fine two-stage formulation that restructures action generation into a coarse initialization step that constructs an action-aware starting point, followed by a single-step local refinement that corrects residual errors. Concretely, the coarse stage learns a conditional posterior over endpoint velocity to transform Gaussian noise into a structured initialization, while the fine stage performs a fixed-time refinement from this initialization. To stabilize training, we introduce a stepwise strategy that first learns a controlled coarse predictor and then performs joint optimization. Experiments on CALVIN and LIBERO show that our method establishes a strong efficiency-performance frontier under low-NFE (Number of Function Evaluations) regimes: it consistently outperforms existing NFE=2 methods, matches or surpasses the NFE=10

π_{0.5}

baseline on several metrics, reduces action sampling latency by 75.4%, and achieves the best average real-robot success rate of 83.0%, outperforming MIP by 19.5 points and

π_{0.5}

by 4.0 points. These results suggest that structured, coarse-to-fine generation enables both strong performance and efficient inference. Our code is available at https://github.com/EmbodiedAI-RoboTron/CF-VLA.

Abstract

Explore similar work

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference

VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting