Organizations: School of Computer Science, Shanghai Jiao Tong University Shanghai, China · Computer Science, King Abdullah University of Science and Technology Thuwal, Mecca, Saudi Arabia
Vision-Language-Action (VLA) models have demonstrated remarkable capabilities in the field of embodied AI, but their high computational cost and limited predicted action length hinder real-time deployment. Although Dadu-Corki, a dedicated accelerator for efficient embodied AI, has been introduced, it does not exploit the inherent interaction patterns between the robot and its environment, which results in a relatively short predicted action length. We observe that robotic environments naturally alternate between active states-where precise actions are crucial-and inactive states-where actions have limited impact on task success. This insight enables a new scheduling opportunity: long-action-length speculative prediction in inactive states, paired with selective verification in active states. We propose SpecVLA, an algorithm-system co-design framework that adaptively balances action length, inference latency, and task reliability. On the algorithm side, SpecVLA introduces a state-aware VLA inference execution paradigm and a hardware-friendly construction of a smaller verification model (sVLA) using differential residuals and block-wise mixed-precision quantization. On the system side, we develop a heterogeneous architecture consisting of a GPU and a robotic-specific hardware module, along with a speculative dataflow that decouples VLA and sVLA through parallel execution. Comprehensive evaluations on OpenVLA and RDT across LIBERO and ManiSkill benchmarks show that SpecVLA reduces end-to-end latency significantly while preserving task success rate. By enabling long-action-length speculative prediction with timely verification, SpecVLA achieves real-time robotic manipulation with both high efficiency and reliability.
Figures & tables
Figure 1. Overview of SpecVLA: (a) A task example of a robot; (b) Existing VLA algorithm; (c) SpecVLA paradigm.
Figure 2. The execution process of the VLA model.
Figure 3. Data distribution of (a) raw features, and (b) residual features.
Figure 4. Success rate of VLA and sVLA under different action length.
Figure 5. Explore environmental state: success rate under different noise patterns.
Figure 6. The latency of SpecVLA and baseline VLA model on A100 GPU.
Figure 7. The execution process of SpecVLA.
Figure 8. The process of the differential-based, hardware-friendly quantization.
Figure 9. Details of SpecVLA dataflow: (a) Sequential dataflow; (b) Speculative dataflow with homogeneous gpu support; (c) Speculative dataflow with heterogeneous architecture support.
Figure 10. The overview of the SpecVLA heterogeneous architecture.
Figure 11. The structure of the state predictor.
Figure 12. The structure of the preprocessing module.
Benchmark
OpenVLA-Spatial
OpenVLA-Object
OpenVLA-Goal
OpenVLA-10
RDT-PickCube
RDT-PushCube
RDT-PlugCharger
State Proportion (%)
Active State
24.59
24.12
22.57
27.76
26.19
26.64
28.48
Inactive State
75.41
75.88
77.43
72.24
73.81
73.36
71.52
Rollback Frequency (%)
All Task
8.43
7.33
8.21
9.45
8.78
9.44
15.23
Complex Task
12.57
12.27
13.17
14.16
11.82
12.58
16.41
Effective Action Length
Spec-VLA ( Wang et al., 2025a )
2.94
2.38
2.14
2.10
-
-
-
HeiSD ( Zheng et al., 2026 )
4.75
4.94
4.83
4.96
-
-
-
Table 1. Quantitative Analysis of SpecVLA Algorithm across Different Benchmarks and Metrics.
Figure 13. Model success rate results: (a) Performance under noise-free environments; (b) Performance with incorporated sensor noise.
Figure 14. The proportion of zero block, low-precision block and high-precision block in the sVLA model.
Method
Setting
Success Rate (%)
FLOPs (%)
Speedup
Spatial
Object
Goal
10
OpenVLA
FP16 (GPU)
98
98
97
93
100
1
QVLA ( Xu et al., 2026b )
W4A8 (GPU)
95
96
94
90
12.63
1.47
QuantVLA ( Zhang et al., 2026a )
W4A8 (GPU)
89
94
90
87
16.13
1.32
Eventful ( Dutson et al., 2023 )
Pruning (GPU)
93
92
94
89
56.38
2.54
MaskVD ( Sarkar et al., 2025 )
Pruning (GPU)
94
95
93
89
69.98
1.76
Table 2. Quantitative Comparison of Differential-based, Hardware-friendly Quantization with Existing Quantization Methods and Pruning Methods on LIBERO Benchmark.
Figure 15. Per-action latency results.
Method
Setting
Success Rate (%)
Speedup
Energy Efficiency
Spatial
Object
Goal
10
OpenVLA
FP16 (GPU)
98
98
97
93
1
1
MoLe-VLA ( Zhang et al., 2026b )
Skipping Layer (GPU)
97
98
96
92
1.98
1.98
VLA-Cache ( Xu et al., 2026a )
Token Reusing (GPU)
97
96
96
92
1.26
1.26
Spec-VLA ( Wang et al., 2025a )
Speculation (GPU)
94
94
96
87
1.32
1.32
HeiSD ( Zheng et al., 2026 )
Speculation (GPU+CPU)
94
98
92
86
2.04
1.69
Table 3. Quantitative Comparison of SpecVLA with Existing Algorithm-only Acceleration Works on LIBERO Benchmark.
Modules
Components
Area( mm2 )
Power( mW )
PE Array
8192 4-bit × 16-bit PE Units
1.58
495.06
680KB On-chip Buffer
State Predictor
7 Subtractors
0.03
13.42
3-stage Adder Trees
Preprocessing Module
8 × 32 Subtractors
0.32
59.26
8 × 32 Multipliers
Table 4. Area of the Robotic-specific Hardware Module.
Figure 16. Energy efficiency results.
Figure 17. Ablation study of software and hardware contributions. Configurations v1–v5 correspond to: (1) Traditional VLA on GPU; (2) sequential SpecVLA (speculative action length = 8) on GPU; (3) sequential SpecVLA (speculative action length = 8) with verification on GPU; (4) SpecVLA with speculative dataflow on GPU; (5) SpecVLA with speculative dataflow and the robotic-specific hardware module.
School of Computer Science and Engineering, Southeast University, Nanjing, China · Nanyang Technological University, Singapore · School of Cyber Science and Engineering, Southeast University, Nanjing, China +1