Barrier-Constrained Guidance For Real-time Asynchronous VLA Execution
Authors: Wenbo Yu, Jifan Li, Wenke Xia, Yuchen Li, Yizhuo Zhang, Pei Ren, Zhengping Che, Di hu
Organizations: Beijing Forestry University · Beijing Innovation Center of Humanoid Robotics · Gaoling School of Artificial Intelligence, Renmin University of China
While asynchronous execution enables VLA models to operate in real-time environments by predicting the next chunk during execution, it inevitably introduces boundary misalignment with the committed action stream. Existing delay-aware solutions address this at the prohibitive cost of extensive retraining, architectural modifications, or fixed-prefix inpainting that inflates computation when low latency is vital. To address these problems, we propose FreeBar, a training-free real-time inference plugin that enables robust asynchronous execution with only marginal additional inference cost by casting chunk alignment in flow-based VLAs as barrier-guided action-space constraint. Specifically, FreeBar introduces a lightweight Control Barrier Function (CBF)-inspired consistency constraint between consecutive action chunks and solves the resulting correction in closed form, avoiding costly backpropagation through the VLA action head. By progressively tightening the consistency bound during generation, FreeBar aligns generated chunks with the committed execution stream while preserving early-stage generative and within-chunk future action flexibility,. Across Kinetix, RoboCasa, and diverse real-world robotic platforms with multiple VLA backbones, FreeBar improves success rates over naive asynchronous execution by 10%--15% on average, with larger gains in highly dynamic scenarios. It also preserves real-time computational efficiency, decreasing inference latency per action chunk over 45% than RTC. Additionally, as a universal plugin, FreeBar remains robust across distinct latency environments, base policies, and hyperparameter settings.
Figures & tables
Figure 1: FreeBar enables smooth, efficient, and robust asynchronous VLA execution. (Top) Swift and fluent robotic responses to dynamic targets (e.g., a moving button). (Bottom) FreeBar ensures smoother kinematics than RTC, efficient inference via lightweight constraint over costly VJP, and robust stability across diverse latencies, frameworks, and hyperparameters. ( ∗ GR00T on RoboCasa, 4 denoising steps.)
Figure 2: Progressive barrier-guided chunk alignment in FreeBar. During asynchronous execution, FreeBar aligns the newly generated action chunk with a time-aligned reference from the previous chunk using closed-form action-space projection. The barrier constraint is progressively tightened across denoising steps and graded along the horizon: proximal actions are strongly aligned, medial actions are softly regularized, and distal actions remain unconstrained.
λ←max(∥g∥22+ϵbk−g⊤vk,0)
Algorithm 1 FreeBar Inference
Figure 3: Left : Success rates across varying execution horizons, with the inference delay fixed at 1. FreeBar consistently outperforms RTC and yields substantial improvements over the naive baseline. Right : Success rates across varying inference delays, where the execution horizon is set to h=max(d,1) , with d being the delay. FreeBar maintains superior performance over RTC under high-latency conditions. Each data point is averaged over 100 trials.
Base Policy: GR00T
Base Policy: π0
Method
d=1
d=2
d=3
d=1
d=2
d=3
Naive
68.0
69.6
68.6
33.8
32.7
31.2
RTC
69.6
70.5
70.0
31.3
32.7
31.7
CBF
68.5
69.8
68.0
31.3
33.3
34.0
FreeBar
72.8
72.3
71.4
36.0
36.5
36.3
Table 1: Left: RoboCasa results (%) with GR00T and π0 as base policies. Right: Hyperparameter robustness across tested configurations, where higher mean performance and lower standard deviation indicate stronger and more stable robustness.
Figure 4: Left : Simulation inference time on RoboCasa with GR00T. FreeBar incurs minimal overhead, comparable to non-horizon-wise CBF, while significantly outperforming RTC. Right : Real-world inference time on the conveyor-belt task with π0 and π0.5 . FreeBar maintains substantially lower latency than RTC.
Benchmark
Method
d=1
d=2
d=3
Kinetix
Temporal Ensembling
41.8
17.2
3.3
FreeBar
87.2
84.5
78.0
RoboCasa
Action-Space Smoothing
70.7
68.9
69.7
Chunk-Boundary Interpolation
65.5
64.2
69.7
Uniform Hard Alignment
69.3
68.9
67.9
FreeBar
72.8
72.3
71.4
Table 2: Comparison with lightweight temporal-consistency baselines. FreeBar consistently outperforms generic smoothing and alignment strategies under varying inference delays d .
Figure 5: Real-world kinematic smoothness comparison between RTC and FreeBar. Top row : Conveyor-belt manipulation (TianGong). Bottom row : Whack-a-mole (Franka Emika robot). Middle & Right : Wrist joint velocity and acceleration. During active manipulation, RTC exhibits remarkable oscillations in both metrics, whereas FreeBar yields substantially smoother profiles.
Method
Conveyor Belt ( π0 )
Conveyor Belt ( π0.5 )
Whack-a-Mole ( π0 )
h=6
h=8
h=10
h=12
h=6
h=8
h=10
h=12
h=6
h=8
h=10
h=12
Naive
56.67
70.00
66.67
63.33
63.33
66.67
76.67
63.33
2.03
5.02
7.05
5.98
RTC
73.33
83.33
80.00
76.67
76.67
80.00
83.33
80.00
10.05
9.98
10.02
8.01
Ours
90.00
86.67
86.67
76.67
93.33
90.00
90.00
83.33
9.61
13.04
12.58
10.62
Table 3: Real-world manipulation results at a fixed inference delay ( d=2 ) across varied execution horizons ( h ). We report Conveyor-Belt Manipulation success rates (%) and total Whack-a-Mole hits within 30 s. Our method achieves the highest performance across almost all tested settings.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Suite
# Tasks
Embodiment / Policy
Domain
Main Challenge
Kinetix
12
Flow policy with action chunking
Physics-based simulation
Fast reaction, contact-rich dynamics, delayed execution
Table A1: Summary of task suites used in our evaluation. The selected tasks cover diverse embodiments, visual domains, control frequencies, and dynamic interaction patterns.
Figure A1: Representative Kinetix tasks from our evaluation suite. The six environments shown cover classic balance control (a, b), reactive manipulation (c, d), and dynamic projectile/locomotion tasks (e, f). In each task the agent must guide a green object to touch a blue target without contacting red obstacles using force-based control. See Table A2 for the full list of all 12 tasks.
ID
Task
Dynamic property evaluated
1
car_launch
Launching a vehicle toward a target; requires precise force magnitude and timing under delayed execution.
2
cartpole_thrust
Balancing an inverted pendulum with active thrust; demands continuous real-time correction to maintain stability.
3
catapult
Catapulting an object to a target; actions have irreversible delayed consequences once the projectile is released.
4
catcher_v3
Catching a moving object mid-flight; requires rapid reaction to a moving target whose position changes during inference.
5
chain_lander
Landing a chain-linked articulated body; coordinated joint control under kinematic constraints and gravity.
6
grasp_easy
Grasping an object in a dynamic scene; requires stable interaction as the object state evolves during inference.
Appendix
Table A2: Description of the 12 Kinetix tasks used in our evaluation.
Figure A2: Visualization of the 11 RoboCasa tasks used in our evaluation. The tasks cover visually rich household scenes and robot manipulation, complementing the highly dynamic physics-control tasks in Kinetix.
ID
Task
Manipulation property evaluated
1
CloseFridge
Pushing a fridge door shut; requires precise contact force and smooth deceleration at chunk boundaries to avoid bounce-back.
2
CloseToasterOvenDoor
Closing a small articulated door; constrained joint motion amplifies boundary discontinuity under delayed execution.
3
CoffeeSetupMug
Placing a mug under a coffee machine; multi-step alignment task where early chunk errors propagate to later stages.
4
NavigateKitchen
Mobile-base navigation in a cluttered kitchen; long-horizon trajectory where stale observations cause spatial drift.
5
OpenCabinet
Pulling open a cabinet door; requires stable chunk transitions during the grasp-then-pull sequence.
6
OpenDrawer
Sliding a drawer open; linear constrained motion where jitter at chunk boundaries can jam the drawer track.
Appendix
Table A3: Description of the 11 RoboCasa tasks used in our evaluation.
Figure A3: Top :Conveyor-Belt Manipulation Bottom :Whack-a-Mole Real-world dynamic manipulation tasks. Conveyor-Belt Manipulation evaluates dual-arm sorting of moving objects, while Whack-a-Mole evaluates single-arm reactive control under abrupt target changes. Both tasks expose severe state-action mismatch under asynchronous inference.
Task
Robot platform
Evaluation metric
Main dynamic challenge
Conveyor-Belt Manipulation
Dual-arm humanoid robot station
Sorting success rate over 30 rollouts
Target objects move during inference, so stale observations lead to grasp-location mismatch and sorting failure.
Whack-a-Mole
Franka Emika 7-DOF arm
Number of successful hits within 30 seconds
Active targets change abruptly, requiring rapid reaction while maintaining smooth chunk transitions.
Appendix
Table A4: Detailed description of the real-world tasks. Both tasks require real-time execution under non-stationary scene dynamics.
Method
0.1
0.5
1
5
10
30
50
Mean ± Std.
RTC
68.0
66.6
–
70.0
69.6
68.6
52.8
65.93±6.55
FreeBar
70.6
–
71.4
70.4
70.6
67.6
67.2
69.63±1.77
Appendix
Table A5: Complete hyperparameter robustness results on RoboCasa simulation. Each entry reports the average task success rate over the evaluated RoboCasa tasks.
Method
0.1
0.5
1
3
5
8
10
Mean ± Std.
RTC
53.33
56.67
67.67
53.33
–
46.67
–
55.53±7.70
FreeBar
66.70
–
73.30
76.67
83.33
–
73.33
74.67±6.04
Appendix
Table A6: Complete hyperparameter robustness results on the real-world Conveyor-Belt Manipulation task with π0 . Each entry reports the sorting success rate (%).
Method
0.1
0.5
1
3
5
8
10
Mean ± Std.
RTC
53.33
63.33
73.33
56.67
–
50.00
–
59.33±9.25
FreeBar
70.00
–
73.30
80.00
80.00
–
76.67
75.99±4.35
Appendix
Table A7: Complete hyperparameter robustness results on the real-world Conveyor-Belt Manipulation task with π0.5 . Each entry reports the sorting success rate (%).
Method
0.1
0.5
1
3
5
8
10
Mean ± Std.
RTC
5.01
6.95
10.08
8.95
–
8.04
–
7.80±1.92
FreeBar
10.06
–
12.03
9.95
11.02
–
8.97
10.40±1.14
Appendix
Table A8: Complete hyperparameter robustness results on the real-world Whack-a-Mole task with π0 . Each entry reports the number of successful hits within 30 seconds.
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.
Real-time execution is crucial for deploying Vision-Language-Action (VLA) models in the physical world. Existing asynchronous inference methods primarily optimize trajectory smoothness, but neglect the critical latency in reacting to environmental changes. By rethinking the notion of reaction in action chunking policies, this paper presents a systematic analysis of the factors governing reaction time. We show that reaction time follows a uniform distribution determined jointly by the Time to First Action (TTFA) and the execution horizon. Moreover, we reveal that the standard practice of applying a constant schedule in flow-based VLAs can be inefficient and forces the system to complete all sampling steps before any movement can start, forming the bottleneck in reaction latency. To overcome this issue, we propose Fast Action Sampling for ImmediaTE Reaction (FASTER). By introducing a Horizon-Aware Schedule, FASTER adaptively prioritizes near-term actions during flow sampling, compressing the denoising of the immediate reaction by tenfold (e.g., in π0.5 and X-VLA) into a single step, while preserving the quality of long-horizon trajectory. Coupled with a streaming client-server pipeline, FASTER substantially reduces the effective reaction latency on real robots, especially when deployed on consumer-grade GPUs. Real-world experiments, including a highly dynamic table tennis task, prove that FASTER unlocks substantially improved real-time responsiveness for generalist policies, enabling rapid generation of accurate and smooth trajectories.
Vision-Language-Action (VLA) policies increasingly rely on asynchronous inference to hide model latency behind ongoing robot motion. During inference, the robot continues executing its current action chunk while the next chunk is generated from the observation available at inference start. By the time the new chunk is executed, the robot or scene may have changed, so an action that fits the prediction-time state can be misaligned with the execution-time state. Training-time methods modify the policy input or supervised target based on the expected execution delay, but do not directly supervise how actions should change as the scene evolves between prediction and execution. Runtime repair methods improve chunk continuity or modify actions during deployment, but can still leave actions aligned with the prediction-time state rather than the execution-time state. Our key insight is that execution-time observations are unavailable when the online prediction is made but are available in recorded trajectories during offline training, making the prediction-execution mismatch observable during training. We propose DEFLECT, an offline post-training framework that turns this temporal contrast into preference supervision for a policy acting from stale inputs. Across Kinetix, LIBERO, and three real-robot tasks, DEFLECT improves delay robustness across the evaluated VLA policies and settings. At the longest supported delay, it improves success by 4.2 percentage points over VLASH on Kinetix and by 22.7 points over naive asynchronous execution with GR00T N1.7 on LIBERO.
Yixiang Zhu, Yonghao Chen, Jiaxiang Zou +2
The Hong Kong University of Science and Technology (Guangzhou) · One Robotics