Vision-Language-Action (VLA) models have demonstrated effectiveness in robot manipulation, yet state-of-the-art models such as pi0.5 operate under a single-frame paradigm, limiting their ability to retain past observations and develop precise spatial perception. In this paper, we propose StreamPI, a streaming multimodal temporal modeling framework that equips single-frame VLA with temporal reasoning capability without introducing any additional parameters. One core design is instruction-anchored temporal modeling. It treats each (visual observation, language instruction) pair as an atomic temporal unit: bidirectional attention within each pair enables cross-modal fusion, while causal attention across pairs preserves autoregressive streaming inference. This ensures the language instruction serves as a persistent semantic anchor throughout task execution. To bridge the gap between synchronous training and asynchronous real-robot deployment, we introduce a andom-interval streaming training strategy: a proper inter-frame interval (e.g., every 3 frames) enables faster and smoother action execution. Beyond this, randomizing the interval further improves robustness to frame-timing perturbations, supporting asynchronous deployment in practice. Furthermore, by leveraging the length extrapolation capability of the LLM backbone, StreamPI seamlessly inherits pretrained single-frame weights and supports flexible single-frame and multi-frame inference. Experiments on real-robot tasks spanning memory-dependent and precise perception scenarios, as well as the simulation benchmark LIBERO, demonstrate that StreamPI outperforms pi0.5 across diverse tasks.
Figures & tables
Figure 1: Comparison of three paradigms for VLA-based robot manipulation. (a) Single-frame VLA: Only the current observation is processed, lacking historical context and limiting temporal memory and spatial perception. (b) Window-based VLA: A window of K frames is processed simultaneously, enriching temporal context at the cost of high computational overhead. (c) Streaming-based VLA (Ours): The model queries cached Key&Value representations from previous timesteps via a lightweight KV cache, achieving temporal memory and precise spatial perception.
Figure 2: The pipeline of StreamPI. To fully unleash the potential of multi-modal interaction, we use bidirectional attention for image-text pairs and causal attention for inter-frames with block-wise causal attention mask. We additional use the random interval sampling to improve the temporal robustness in the real-world deployment.
Figure 3: Visualization of real-world tasks and comparison of real-world performance. Left : Visualization of memory-dependent and precise perception-dependent real-world tasks. Right : Performance comparison of real-world tasks between π0.5 and StreamPI.
Method
Libero-Spatial
Libero-Object
Libero-Goal
Libero-Long
Avg. Success Rate
Diffusion Policy Chi et al. (2025)
78.3
92.5
68.3
50.5
72.4
Octo Ghosh et al. (2024)
78.9
85.7
84.6
51.1
75.1
SpatialVLA Qu et al. (2025)
88.2
89.9
78.6
55.5
71.7
TraceVLA Zheng et al. (2024)
84.6
85.2
75.1
54.1
74.8
OpenVLA Kim et al. (2024)
84.7
88.4
79.2
53.7
75.9
CoT-VLA Zhao et al. (2025)
87.5
91.6
87.6
69.0
81.1
Table 1: Performance comparison on LIBERO ( Liu et al., 2023 ) . Success rates (%) are reported across four suites.
Method
1
2
3
4
5
Avg. Seq. Len
MemoryVLA Shi et al. (2025)
94.8
87.4
81.4
75.9
69.4
4.090
π0.5 Intelligence et al. (2025)
94.2
88.7
85.7
83.2
79.5
4.313
StreamPI ( T=5 )
96.9
93.6
90.7
88.5
85.0
4.547
Table 2: Performance comparison on CALVIN ( Mees et al., 2022 ) . Success rates (%) at each sequence position and average sequence length are reported.
T (Streaming Frames)
Intra-Pair
Inter-Pair
Spatial
Object
Goal
Long
Avg. Success Rate
1
Causal
—
97.4
97.8
96.8
91.0
95.8
Bidirectional
—
98.4
98.2
97.4
92.0
96.5
3
Causal
Causal
98.6
96.0
95.4
91.0
95.3
Bidirectional
Causal
98.8
98.6
98.6
93.8
97.5
5
Causal
Causal
98.0
97.6
96.8
90.6
95.5
Bidirectional
Causal
98.8
99.8
99.6
95.0
98.3
Table 3: Ablation on attention direction. Intra-pair denotes attention between image and text tokens within each temporal unit. Inter-pair denotes attention across temporal units and is always causal. Results are average success rates (%) on the LIBERO benchmark.
Training δ
Training T
Test T
Spatial
Object
Goal
Long
Avg. Success Rate
π0.5 ( δ=1 )
1
1
98.4
98.2
97.4
92.0
96.5
Fixed ( δ=1 )
3
3
98.0
98.4
96.8
92.2
96.4
5
5
98.4
97.4
98.6
93.4
97.0
Random δ
3
3
98.8
98.6
98.6
93.8
97.5
5
5
98.8
99.8
99.6
95.0
98.3
Table 4: Effect of random-interval streaming training. Models are trained with fixed ( δ=1 ) or random interval ( δ∼U[3,7] ) and evaluated at δ∈{1,3,5} .
Training T
Test T
Spatial
Object
Goal
Long
Avg. Success Rate
π0.5 (baseline)
1
98.4
98.2
97.4
92.0
96.5
T=3
1
98.0
98.8
97.8
93.4
97.0
3
98.8
98.6
98.6
93.8
97.5
T=5
1
97.4
98.8
98.4
93.6
97.1
3
99.0
98.8
98.0
93.8
97.4
5
98.8
99.8
99.6
95.0
98.3
Table 5: Cross-stream generalization. A model trained with T=5 is evaluated at T∈{1,3,5} . Results are average success rates (%) on LIBERO-Long.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Streaming Frames
Inference Time (ms)
1
94.4±3.4
3
97.9±5.1
5
103.6±6.3
8
110.9±10.2
10
117.9±16.5
Appendix
Table 6: The inference time of different streaming frames in streaming inference.
Figure 4: AgileX PiperX robotic arms.
Model
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
S.R. (%)
π0.5
1
0
1
0
0
1
1
1
1
0
1
0
0
0
0
46.7%
StreamPI
1
1
1
0
1
1
1
1
1
1
1
0
1
1
0
80.0%
Appendix
Table 7: Performance comparison of each trial on “Shell Game” task. S.R. denotes success rate.
Model
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
S.R. (%)
π0.5
1
0
1
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
1
0
0
1
0
1
0
1
0
26.7
StreamPI
0
1
1
0
1
0
1
0
0
1
0
1
1
0
1
0
1
1
1
1
1
0
1
1
1
1
0
0
1
1
63.3
Appendix
Table 8: Performance comparison of each trial on “Pick up the Rolling Bottle” task. S.R. denotes success rate.
M-F
M-M
F-M
S.R. (%)
Model
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
π0.5
0
0
0
0
0
0
0
0
0
0
1
0
1
1
1
1
0
1
0
1
0
1
0
1
0
0
1
1
1
0
40.0
StreamPI
1
1
0
1
0
0
1
0
0
1
1
1
1
0
1
0
0
1
1
1
1
1
0
1
1
1
1
0
1
0
66.7
Appendix
Table 9: Performance comparison of each trial on “Pen Insertion into Narrow Bottle” task. S.R. denotes success rate.
Model
Center
Left-Far
Right-Far
Right-Near
Left-Near
S.R. (%)
π0.5
1
0
1
1
1
0
0
0
1
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
60.0
StreamPI
1
1
1
1
1
1
1
1
0
1
1
1
0
1
1
1
1
1
1
1
1
1
1
1
1
92.0
Appendix
Table 10: Performance comparison of each trial on “Cup Insertion into Cup Sleeve” task. S.R. denotes success rate.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
L → M
L → R
L → M → R
L → M → L
L → R → M
M → L
M → L → M
M → R
M → R → M
M → L → R
R → M
R → M → R
R → L
R → L → M
R → L → R
Appendix
Table 11: Each pattern of the Shell Game. ‘L’, ‘M’, and ‘R’ is the left, middle, and right position, respectively. A → B denotes that move the cup that contains the target object from A to B.
Figure 5: Comparison of “Cup Insertion into Cup Sleeve”.
Figure 6: Comparison of “Pen Insertion into Narrow Bottle”.
Figure 7: Comparison of “Pick up the rolling bottle”.
Figure 8: The continuous qualitative results of “Shell Game”. For convenience, we use a red circle to indicate which cup contains the object.
Vision-language-action (VLA) models have demonstrated exceptional performance in natural language-driven perception and control. However, the high computational cost of VLA models poses significant efficiency challenges, particularly for resource-constrained edge platforms in real-world deployments. However, since different stages of VLA (observation, action generation and execution) must proceed sequentially, and wait for the completion of the preceding stage, the system suffers from frequent halting and high latency. To address this, We conduct a systematic analysis to identify the challenges for fast and fluent generation, and propose enabling VLAs with the ability to asynchronously parallelize across VLA stages in a "streaming" manner. First, we eliminate the reliance on action chunking and adopt action flow matching, which learns the trajectory of action flows rather than denoising chunk-wise actions. It overlaps the latency of action generation and execution. Second, we design an action saliency-aware adaptive observation mechanism, thereby overlapping the latency of execution and observation. Without sacrificing performance, StreamingVLA achieves substantial speedup and improves the fluency of execution. It achieves a 2.4 × latency speedup and reduces execution halting by 6.5 ×.
Vision-Language-Action (VLA) models are increasingly promising for robotic manipulation, yet their real-world deployment remains bottlenecked by high inference latency and unstable asynchronous execution. This challenge is particularly pronounced in flow-matching-based VLA models, where action decoding requires multiple iterative steps conditioned on the VLM context. While efficient inference methods improve control frequency and asynchronous methods reduce execution idle time, existing approaches often fail to jointly achieve low-latency inference and accurate, temporally consistent asynchronous execution. We introduce FlashVLA, a streaming action decoding framework that addresses both challenges in a unified formulation. FlashVLA maintains a streaming action buffer with multiple chunks at different noise levels and decodes them using chunk-wise causal attention. This design allows FlashVLA to produce one executable action chunk per inference step. Moreover, its chunk-wise autoregressive formulation implicitly preserves action continuity, enabling smooth asynchronous execution without extra future-state conditioning. Across extensive simulated and real-world experiments, FlashVLA substantially improves inference speed while maintaining strong task performance. It can achieve ≥30,Hz control frequency on a single GPU with smooth asynchronous inference in real-world deployment.
Vision-language-action (VLA) models have achieved great success on general robotic tasks, but still face challenges in fine-grained spatiotemporal manipulation. Typically, existing methods mainly embed spatiotemporal knowledge into visual and action representations, and directly perform a cross-modal mapping for step-level action prediction. However, such spatiotemporal reasoning remains largely implicit, making it difficult to handle multiple sequential behaviors with explicit spatiotemporal boundaries. In this work, we propose ST-π, a structured spatiotemporal VLA model for robotic manipulation. Our model is guided by two key designs: 1) Spatiotemporal VLM. We encode 4D observations and task instructions into latent spaces, and feed them into the LLM to generate a sequence of causally ordered chunk-level action prompts consisting of sub-tasks, spatial grounding and temporal grounding. 2) Spatiotemporal action expert. Conditioned on chunk-level action prompts, we design a structured dual-generator guidance to jointly model spatial dependencies and temporal causality, thus predicting step-level action parameters. Within this structured framework, the VLM explicitly plans global spatiotemporal behavior, and the action expert further refines local spatiotemporal control. In addition, we propose a real-world robotic dataset with structured spatiotemporal annotations for fine-tuning. Extensive experiments have been conducted to demonstrate the effectiveness of our model. Our code link: https://github.com/chuanhaoma/ST-pi.
Chuanhao Ma, Hanyu Zhou, Shihan Peng +3
School of Artificial Intelligence and Automation, Huazhong University of Science and Technology · School of Computing, National University of Singapore · School of Computing, Macquarie University