Deploying Vision-Language-Action (VLA) models on resource-constrained edge platforms encounters a fundamental conflict between high-latency semantic inference and the high-frequency control required for dynamic manipulation. To address the challenge, this paper presents Agile-VLA, a hierarchical framework designed for industrial pose reorientation tasks on edge devices such as the NVIDIA Jetson Orin Nano. The core innovation is an Implicit Affordance Anchoring mechanism that directly maps geometric visual cues, specifically centroid and rim keypoint anchors, into structured parametric action primitives, thereby substantially reducing reliance on high-latency semantic inference during closed-loop control. By decoupling perception (10 Hz) from control (50 Hz) via an asynchronous dual-stream architecture, the system effectively mitigates the frequency mismatch inherent in edge-based robot learning. Experimental results on a standard 6-DoF manipulator demonstrate that Agile-VLA achieves robust rectification of complex, irregular workpieces using only 5-shot demonstrations through extrinsic dexterity.
Figures & tables
Fig. 1: Visualization of Agile-VLA task primitives across diverse industrial workpieces. The top row illustrates Stable Pick and Place operations guided by Stability Anchoring, where the grasp point is optimized toward the geometric centroid to minimize gravitational torque. The bottom row demonstrates Pick, Flip and Place sequences leveraging Pivot Anchoring; here, the anchor point is constrained to the object boundary to maximize the lever arm for extrinsic dexterity. The framework is validated on four irregular objects: (a) Battery, (b) Calculator, (c) Phone, and (d) Chip.
Fig. 2: System Overview. Compared to conventional large-scale VLA models that suffer from high cloud latency, Agile-VLA achieves real-time, few-shot pose reorientation directly on an edge device (NVIDIA Jetson Orin Nano). By decoupling high-level semantic reasoning from low-level control, the framework enables agile manipulation of complex industrial components using only a minimal number of demonstration samples.
Fig. 3: Asynchronous Dual-Stream Framework. The semantic perception stream (10 Hz) is responsible for extracting manipulation anchors, while the proprioceptive control stream (50 Hz) executes parameterized action primitives. A timestamp-based soft synchronization protocol ensures strict alignment between the visual and kinesthetic streams, thereby enabling high-fidelity few-shot adaptation on resource-constrained hardware.
Fig. 4: Implicit Affordance Anchoring Protocol. The system maps visual cues to specific action primitives according to the object’s topological state. Front-facing objects are anchored at their geometric centroid to trigger stable, guidance-oriented grasping, whereas back-facing objects are anchored at the rim to initiate extrinsic-dexterity-based flipping, thereby effectively leveraging environmental constraints to overcome kinematic limitations.
Category
Method
Tinf (ms) ↓
fctrl (Hz) ↑
VRAM (GB) ↓
TCP Jitter (mm) ↓
Sequence
RT-1
125±12
8
3.8
3.25
Diffusion
Octo-Base
245±18
4
3.2
4.12
LLM-VLA
OpenVLA (4-bit)
835±42
1.2
6.8
5.82
Ours
Agile-VLA (Sync)
92±6
10.8
1.8
1.45
Ours
Agile-VLA (Async)
92 (Vis)
50
2.1
0.65
TABLE I: System profiling on Jetson Orin Nano (8GB).
Method
Easy SR (%) ↑
Hard SR (%) ↑
Avg SR (%) ↑
Time (s) ↓
Collision (%) ↓
Jerk (m/s 3 ) ↓
Transporter
74.5
28.6
55.4
11.2
16.5
10.21
RT-1
78.2
34.0
60.1
14.5
18.2
12.84
RVT
82.5
36.5
63.8
12.8
14.2
11.56
Octo-Base
85.0
38.5
61.7
16.8
15.4
14.35
OpenVLA (4-bit)
88.6
46.5
71.0
19.5
21.8
15.62
Agile-VLA
96.5
84.6
90.5
8.4
3.2
1.24
TABLE II: Multi-dimensional performance on DID-127 dataset.
Fig. 5: Real-world multi-task reorientation performance across extreme physical conditions. We systematically benchmark Agile-VLA against state-of-the-art foundation models on six extreme industrial parts (involving heavy masses and complex topologies) and two novel out-of-distribution objects. Crucially , the performance on ’Unseen’ objects reflects the results after our 5-shot edge adaptation pipeline, demonstrating the framework’s rapid deployment capability in contrast to the massive fine-tuning required by large VLA baselines. By leveraging extrinsic pivoting, Agile-VLA successfully bypasses actuator torque overload (e.g., Heavy Battery) and kinematic singularities (e.g., Complex PCB), achieving consistently higher success rates.
Samples ( N )
Pixel Error (px) ↓
Avg SR (%) ↑
Tuning Time (min) ↓
N=1
18.4
35
0.8
N=5
3.5
90
2.5
N=10
3.2
90
5.1
TABLE III: Impact of annotation samples ( N ).
Async
Anchor
Pivot
SR (%) ↑
Jitter (mm) ↓
× (Sync)
✓
✓
60.5
1.45
✓
× (Center)
✓
43.2
0.68
✓
✓
× (Direct)
18.5
0.82
✓
✓
✓
90.5
0.65
TABLE IV: Ablation study on core architecture components.
Vision-Language-Action (VLA) models provide a unified framework for perception, language conditioning, and action generation, but many existing systems remain difficult to deploy in embedded robotic settings because of their computational requirements and inference latency. In this paper, we present LiteVLA-Edge, a deployment-oriented VLA pipeline for fully on-device inference on Jetson Orin-class hardware. Our approach combines supervised image-to-action fine-tuning in FP32 with post-training 4-bit GGUF quantization and GPU-accelerated inference through the llama.cpp runtime. Under our deployment configuration, LiteVLA-Edge achieves a mean end-to-end latency of 150.5,ms (approximately 6.6,Hz) while operating entirely offline within a ROS~2-integrated perception--reasoning--action pipeline. Rather than introducing a new policy objective, our contribution is a practical systems path for executing compact multimodal control models locally on embedded hardware while preserving modular interfaces between perception, reasoning, and actuation. These results establish timing feasibility for reactive language-conditioned control and provide a reproducible baseline for future task-level evaluation of on-device VLAs in robotics.
Justin Williams, Kishor Datta Gupta, Roy George +1
Clark Atlanta University, Atlanta, GA, USA · Siemens Corporation, Princeton, NJ, USA
Vision-Language-Action (VLA) models have achieved impressive performance on diverse embodied tasks. However, deploying VLA models on low-power onboard devices, such as the Jetson Orin, remains challenging due to their high computational complexity, which leads to substantial inference latency and low control frequency. Asynchronous inference can partially mask this latency by parallelizing action execution and subsequent inference, but it introduces two critical issues: perception-execution misalignment and long reaction time. In this paper, we propose EagleVLA, a method for efficient VLA deployment on onboard devices via Foresight-Aligned Asynchronous Correction. To address misalignment, we train a lightweight future correction module that predicts future environment representations, allowing the action expert to predict actions from the future time step. To reduce reaction time, we introduce confidence-based scheduling optimization that adaptively balances VLM and action expert invocations. We also build a llama.cpp-based inference engine tailored for onboard VLA deployment, with system-level optimizations including CUDA graph reuse, GPU-resident intermediate buffering, and flow unrolling. Extensive experiments demonstrate that EagleVLA achieves 9.85x and 6.19x improvements in control frequency compared with naive PyTorch and vla.cpp on Jetson Orin, while outperforming VLASH by 14.8% in success rate on LIBERO benchmark. The code of our asynchronous algorithm is available on https://github.com/PKU-SEC-Lab/EagleVLA, and our efficient llama.cpp-based inference engine is available on https://github.com/PKU-SEC-Lab/EagleVLA-Edge.
Zebin Yang, Qi Wang, Yunhe Wang +6
Peking University · TLAIC · PrimeBot Research Institute +1
Vision-Language-Action (VLA) models have shown strong potential for robotic manipulation, but real-time deployment on edge hardware remains challenging. In this work, we identify VLM visual and context tokens as a major source of deployment latency: for GEMM-dominated projection operators, computation grows linearly with the number of input tokens when model dimensions are fixed. Motivated by this observation, we propose RhinoVLA, a deployment-oriented VLA model co-designed with the Huixi R1 edge SoC. RhinoVLA adopts a token-efficient Qwen3-VL backbone and a continuous Action Expert, reducing the VLM-side token and computation burden while preserving pretrained multimodal capability. To support cross-robot learning, RhinoVLA further introduces a unified interface that combines View Registry, 72D physical state-action slot space, and robotinstance LoRA, allowing heterogeneous robot observations and action schemas to be aligned under a shared policy. On the deployment side, RhinoVLA is optimized through hardware-aware compilation, mixed-precision execution, and parallel visual encoding. Experiments show that RhinoVLA achieves downstream performance comparable to π0.5 at a similar parameter scale, while reaching 11.69 Hz end-to-end inference on Huixi R1, meeting the 10 Hz real-time closedloop control target. The project will be open-sourced at https://github.com/HuixiAI/RhinoVLA.