Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed-accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PaceVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, and (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PaceVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation.
Figures & tables
Figure 1: Adaptive temporal execution with PaceVLA . (a) Conveyor package handling task on the humanoid ALLEX: PaceVLA selects a temporal granularity per phase. (b) Success rate vs. speedup over the base VLA, averaged over six real-world tasks.
Figure 2: Overview of PaceVLA . PaceVLA is a unified framework for adaptive temporal action execution in pretrained VLAs. It equips a shared action expert with mode-specific action decoders that support different temporal granularities and prediction horizons.
Figure 3: Experimental setup. Top: Three simulation benchmarks, LIBERO, RoboCasa-Kitchen, and DexJoCo. Bottom: Real-world manipulation tasks across three embodiments, a Franka Research Arm with a gripper, an OpenArm robot with a dexterous hand, and an ALLEX humanoid. The real-world evaluation includes long-horizon and dexterous manipulation tasks.
Method
RoboCasa-Kitchen
LIBERO
PnP
Others
Short
Long
SR ↑
TPR ↑
ES ↑
SR ↑
TPR ↑
ES ↑
SR ↑
TPR ↑
ES ↑
SR ↑
TPR ↑
ES ↑
RLDX-1-IMG [ 18 ]
48.5
1.10
1.00
76.3
2.73
1.00
96.7
7.93
1.00
88.6
3.15
1.00
+ Fixed Quantization [ 30 ]
43.0
1.55
0.87
72.0
4.00
0.97
89.4
11.80
1.23
79.2
4.44
1.10
+ DemoSpeedUp [ 13 ]
28.7
0.80
0.49
62.7
3.37
0.68
93.5
14.03
1.54
75.2
4.95
1.06
+ SAIL [ 2 ]
38.5
1.03
0.72
66.6
2.80
0.75
96.5
10.34
1.27
76.6
3.38
0.92
Table 1: Simulation benchmark results. We compare PaceVLA with baseline methods on RoboCasa-Kitchen and LIBERO. For RoboCasa-Kitchen, we report results on the pick-and-place tasks (PnP) and the remaining tasks (Others), with models trained using 300 demonstrations per task. For LIBERO, we report results on the short-horizon suites (Short, averaged over Spatial, Object, and Goal) and the long-horizon suite (Long). We report the success rate (SR, %), throughput-with-regret (TPR, ×10−3 ), and the effective speedup (ES). Base models are listed for reference ( ES=1 by definition). Among the acceleration methods for each base model, the best and second-best results are in bold and underlined , respectively.
Franka + Gripper
OpenArm + Bimanual Hand
Pick-and-Place
Air Fryer
Stack Cup
Table Sweeping
Pump and Scrub
Method
SR ↑
Time (s) ↓
SR ↑
Time (s) ↓
SR ↑
Time (s) ↓
SR ↑
Time (s) ↓
SR ↑
Time (s) ↓
GR00T N1.5 [ 5 ]
81.3
12.1
81.3
24.6
84.4
35.3
62.5
21.9
64.1
30.0
+ Fixed Quantization [ 30 ]
62.5
7.7
53.1
14.2
62.5
18.7
48.4
13.3
42.2
19.9
+ DemoSpeedUp [ 13 ]
68.8
8.1
71.9
18.9
53.1
24.2
56.3
17.9
53.1
20.2
+ SAIL [ 2 ]
65.6
7.9
56.3
16.8
45.3
20.0
46.9
15.4
45.3
15.3
Table 2: Real-World evaluation results. We report average success rate (SR, %, over 32 trials per task; partial scores of 0.5 are given on Stack Cup, Table Sweeping, and Pump and Scrub, see Appendix C ) and execution time (Time, average wall-clock execution time on fully successful trials) on five real-world tasks across two embodiments: three Franka+Gripper tasks ( Pick-and-Place, Air Fryer, Stack Cup ) and two OpenArm+Hand tasks ( Table Sweeping, Pump and Scrub ).
Table 5: Ablation study. Success rate (SR, %), throughput-with-regret (TPR, ×10−3 ), and effective speedup (ES) on RoboCasa-Kitchen, averaged over all 24 tasks. ES is computed relative to the base model GR00T N1.5. Our default setting is highlighted in violet .
Method
Latency (ms) ↓
Rel. Cost
GR00T N1.5
40.07
1.00×
PaceVLA
42.15
1.05×
Table 6: Inference cost. Per-step latency on RoboCasa-Kitchen.
Figure 4: Quantizability supervision analysis. We visualize the quantizability scores annotated on a RoboCasa-Kitchen demonstration. Shaded bands indicate manipulation phases. Scores are low around contact transitions (grasp and place) and higher during reach and carry, so coarse execution is suppressed exactly where precise control matters.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Hardware setups and task environments.
GR00T N1.5 + PaceVLA
RLDX-1-IMG + PaceVLA
Optimizer
AdamW
AdamW
Optimizer momentum
β1,β2=0.95,0.999
β1,β2=0.9,0.999
Optimizer weight decay
1e-5
1e-5
Learning rate
1e-4
1e-4
Learning rate scheduler
Cosine decay
Cosine decay
Warmup iterations
3,000
3,000
Appendix
Table 7: Hyperparameter details for training PaceVLA on simulation benchmarks.
Franka
OpenArm
ALLEX
Batch size
64
64
256
Training iterations
30,000
40,000
30,000
Fine-grained modes
(16,16),(8,8)
(16,16),(8,8)
(24,24),(12,12)
Quantized modes
(16,8),(8,4)
(16,8),(8,4)
(24,12),(12,6)
Appendix
Table 9: Embodiment-specific settings for real-world tasks. Modes are given as (Hk,Tk) .
Decoder (H,q)
Metric
Ours
Base
(16,2)
MSE ↓
0.0273
0.0344
DTW ↓
3.209
3.777
Cosine ↑
0.850
0.809
(8,2)
MSE ↓
0.0163
0.0243
DTW ↓
1.359
1.658
Cosine ↑
0.886
0.847
Appendix
Table 10: Quantized-action prediction accuracy on RoboCasa-Kitchen. Ours uses dedicated coarse decoders, while Base quantizes GR00T N1.5 predictions using the same SO(3)-aware rule. (H,q) denotes the source horizon and quantization level, and Both averages the two configurations.
Figure 6: End-effector rollout trajectories. End-effector trajectories of the base policy (blue) and PaceVLA (orange) from the same initial state; PaceVLA traverses the same path with fewer, sparser steps.
Figure 7: Failure cases of PaceVLA . Most failures come from small targeting errors (Pick-and-Place, Stack Cup) or from the hands and tools blocking the contact region in the egocentric view (Table Sweeping, Pump and Scrub).
Robot manipulation alternates between low-risk transit phases that call for fast execution and high-risk contact stages that demand slow, precise motion. Yet existing Vision-Language-Action models (VLAs) only inherit a single fixed speed from training demonstrations. Prior efforts to accelerate VLAs through model compression, KV-cache reuse, or reinforcement learning only shift the policy from one fixed speed to another, and leave deceleration almost unexplored. We observe that the magnitude of each predicted action already governs how fast the robot moves, opening a direct route to controllable execution speed. We turn this observation into TempoVLA, a single VLA whose execution speed is controlled by an explicit condition. TempoVLA combines two coupled components. (1) A data-side Variable-Speed Trajectory Augmentation (VSTA) that re-times demonstration to any target speed by merging or splitting actions while preserving its motion semantics. (2) A model-side conditioning mechanism that feeds the speed to the policy. Statistics show that VSTA reaches the requested speed with negligible motion error. Experiments in simulation and on real-world tasks demonstrate that TempoVLA achieves flexible speed control in both directions, while VSTA additionally boosts the default 1× performance via better data utilization. Furthermore, by cooperating with a large multimodal model, TempoVLA realizes dynamic speed control, accelerating through low-risk phases and decelerating for high-risk ones.
Vision-Language-Action (VLA) models are a powerful paradigm for generalist robotic control. However, their high computational cost and limited control frequency hinder real-time robotic manipulation, especially when large vision-language backbones and iterative action heads run at every control step. Existing VLA acceleration methods often optimize individual components or rely on fixed acceleration rules, treating different control steps with largely fixed computation and overlooking the non-uniform reasoning demands of sequential embodied control. Inspired by human motor control, where cognitive and feedback resources concentrate on goal-sensitive stages, we argue that VLA models should learn when to invest full computation and when to reuse prior computation. We propose ElegantVLA, a plug-in phase-adaptive inference framework that accelerates VLA models through intra-model dynamic compute scheduling. ElegantVLA introduces a lightweight scheduler that observes temporal representation similarity, robot-motion cues, and episode progress to jointly allocate computation across the vision encoder, LLM, and action head. For perception-language reasoning, the scheduler selects a five-level Vision-LLM compute mode, from full recomputation to multi-step temporal reuse, based on visual-language representation stability. For action generation, it selects a three-level denoising mode, reusing intermediate denoising states during stable motion while preserving full refinement for goal-sensitive stages. By coordinating these decisions, ElegantVLA offers a general acceleration framework for modern VLA pipelines with explicit action-generation modules, without modifying or retraining the base model. Experiments on GR00T and CogACT achieve up to 2.55x and 3.77x speedup, and on six real-world GR00T tasks ElegantVLA cuts computation by 2.18x while raising control frequency from 13.8 Hz to 26.3 Hz.
Ye Li, Huanan Liu, Kangye Ji +7
Tsinghua University · University of Illinois at Urbana-Champaign
Vision-Language-Action (VLA) models exhibit strong generalization for robotic manipulation, yet their high inference latency limits real time deployment. We identify two primary sources of temporal redundancy in existing VLA pipelines: repeated visual encoding of highly similar consecutive frames and multi step iterative sampling in diffusion based policies. To address this, we propose a system level acceleration strategy that reduces computation in both perception and action generation. On the perception side, we incrementally update only tokens corresponding to dynamic scene regions instead of re-encoding entire frames. On the policy side, we compress diffusion sampling into a compact 2-step schedule through efficiency oriented training while preserving action precision. Experiments on Libero, RobotWin, and Real Robot Platforms demonstrate over 2 times speedup while maintaining high performance, achieving up to 98% success rate on general manipulation benchmarks. Our codes will be released on Github.
Yuzhou Wu, Yuxin Zheng, Muchun Niu +6
1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2