EcoVLA: Energy-Efficient Device-Edge Co-Inference for Vision-Language-Action Models under Real-Time Constraints

Ao Zhou; Bo Dai; Le Yu; Xingyu Liu; Zeyu Hao; Lingkun Long; Chunming Hu; Jianlei Yang

Source-linked AI summary

EcoVLA: Energy-Efficient Device-Edge Co-Inference for Vision-Language-Action Models under Real-Time Constraints

Ao Zhou, Bo Dai, Le Yu, Xingyu Liu, Zeyu Hao, Lingkun Long, Chunming Hu, Jianlei Yang

arXiv:2608.15502v1Aug 16, 2026cs.AIcs.RO

TL;DR

VLA inference is difficult to deploy on robots because limited device resources conflict with real-time control and energy-efficiency requirements. EcoVLA adaptively co-infers across device and edge using unified stage-level modeling and runtime scheduling, improving energy efficiency by up to 236% over fixed co-inference baselines while satisfying SLOs at 20 Hz.

Problem
VLA models’ computational cost conflicts with robots’ limited compute and energy budgets, hindering real-time deployment and motivating system-level energy optimization.
Method
EcoVLA uses an architecture-agnostic stage-level co-inference abstraction, joint latency-energy modeling, and adaptive SLO-constrained scheduling across devices and edge servers.
Results
Up to 236% higher system energy efficiency than fixed co-inference baselines was achieved under a 20 Hz action output constraint while consistently satisfying the SLO.
Takeaways & Limitations
EcoVLA supports energy-efficient real-time VLA deployment across heterogeneous models and systems under varying runtime conditions.
Takeaways & Limitations
The framework is scoped to device-plus-single-edge deployment, with multi-edge collaboration left for future work.

Abstract

from arXiv · show

Vision-Language-Action (VLA) models have emerged as a promising foundation for Embodied AI, but their high inference cost poses significant challenges for deployment in robotic systems. In practice, on-device inference is constrained by limited compute capacity and energy budgets, struggling to simultaneously satisfy real-time control and energy efficiency requirements. Alternatively, offloading the inference workload to an edge server is susceptible to fluctuations in system conditions, introducing unpredictable latency risks. Device-edge co-inference offers a promising solution, but systematic research tailored to VLA models remains scarce, particularly a unified co-inference framework that jointly addresses real-time constraints and system-level energy efficiency. Thus, we propose EcoVLA, an adaptive device-edge co-inference framework for VLA models that maximizes system energy efficiency under real-time constraints. EcoVLA first introduces a unified stage-level abstraction over different VLA paradigms, establishing an architecture-agnostic co-inference design space. It then formulates a joint device-edge-network latency and energy prediction model to enable rapid runtime evaluation of candidate co-inference schemes. Building on this, EcoVLA continuously selects the energy-optimal scheme satisfying real-time constraints with millisecond-level overhead, adapting to runtime variations in network and system states. Furthermore, EcoVLA incorporates a lightweight transmission mechanism for inter-stage intermediate tensors to reduce the communication overhead incurred by cross-device collaboration. Experimental results across VLA models show that EcoVLA improves system energy efficiency by up to 236% over existing co-inference approaches under a 20 Hz action output frequency constraint, while consistently maintaining SLO satisfaction under dynamic network and edge workload conditions.

1 Introduction

EcoVLA addresses the mismatch between VLA inference demands and robot resource limits by adaptively coordinating device-edge execution under real-time constraints. It maximizes Actions/J through architecture-agnostic staging, runtime prediction, and energy-optimal scheduling.

Introduction: VLA models’ prohibitive computational cost conflicts with robots’ limited compute capability and energy budgets.On-device deployment therefore struggles to meet both efficiency and real-time requirements.
Introduction: Full-model edge offloading lacks latency guarantees under network fluctuations and concurrent access from multiple robots.This makes inference timeouts difficult to avoid.
Introduction: System-level energy efficiency must be evaluated across the entire control cycle, including structural idle time before the next command.Inference speed alone does not determine robotic control-loop efficiency.
Introduction: EcoVLA introduces a paradigm-agnostic adaptive device-edge co-inference framework that maximizes Actions/J under real-time constraints through stage-level execution graphs.The abstraction supports unified design and deployment across heterogeneous VLA models and systems.
Introduction: Over 95% prediction accuracy within a 20% error bound enables dynamic scheduling that selects energy-optimal SLO-feasible configurations.EcoVLA improves system energy efficiency by up to 236% over fixed co-inference baselines under a 20 Hz action output frequency constraint, while consistently satisfying the SLO.

2 Related Work

Prior work identifies autoregressive and diffusion-based VLA paradigms, both constrained by on-device compute and power budgets that hinder real-time robot control. Device-edge co-inference has evolved from layer-level DNN partitioning toward communication-aware, system-level co-design for complex and generative models.

VLA inference efficiency: VLA models primarily use autoregressive or diffusion-based designs, yet both incur inference costs exceeding on-device compute and power budgets.These limitations make satisfying real-time robot-control requirements difficult.
Device-edge co-inference: Device-edge co-inference partitions computation between on-device processors and edge servers to overcome single-device resource limitations.This approach addresses the resource constraints that motivate distributed inference.
Device-edge co-inference: Co-inference research progressed from layer-level partitioning for classical DNNs to communication-aware mapping and system-level performance prediction for GNNs.The GNN direction jointly considers communication and system-level performance.
Device-edge co-inference: More recent co-inference research has expanded to generative models, including mechanisms oriented toward LLMs.The supplied passage identifies this expansion but does not specify the mechanisms further.

3 Motivation and Overview

EcoVLA is motivated by three challenges in VLA device-edge co-inference: heterogeneous architectures, dynamic runtime conditions, and system-level energy efficiency. It addresses them with a unified stage-level design space and joint latency-energy awareness for candidate schemes.

Challenge 1: Heterogeneous VLA paradigms: Heterogeneous autoregressive and diffusion-based VLA paradigms differ in architectures, execution flows, stage boundaries, intermediate tensors, and feasible partition granularity.OpenVLA uses visual encoding, language-conditioned fusion, and step-wise action decoding, while π0 generates trajectories through iterative denoising.
Challenge 2: Dynamic runtime conditions: Dynamic network fluctuations and edge workload variations make pure edge offloading and fixed device-edge partitioning unreliable for sustained real-time VLA inference.For OpenVLA inference, these static strategies cannot reliably satisfy real-time requirements over time.
Challenge 3: System energy efficiency: System energy efficiency cannot be inferred from device power or inference latency alone, making intuition-driven partitioning unreliable under real-time constraints.The deployment objective is system-level energy efficiency rather than minimizing device power in isolation.
EcoVLA overview: EcoVLA decouples heterogeneous VLA paradigms into a common executable graph and mapping space through a unified stage-level co-inference design space.This abstraction is intended to provide a common basis for co-inference across different VLA paradigms.
EcoVLA overview: EcoVLA jointly estimates candidate schemes’ end-to-end latency and energy efficiency by combining offline stage-level profiling with online network and edge information.The framework uses system performance awareness to evaluate candidate co-inference schemes.

4 Unified VLA Co-Inference Design Space

EcoVLA unifies heterogeneous VLA architectures through a stage-level DAG, standardized intermediate-state interface, and device-mapping space. This abstraction supports executable, deployable device-edge co-inference across autoregressive and diffusion-based VLA paradigms.

Unified stage-level abstraction: EcoVLA represents each VLA inference as a directed acyclic graph whose vertices are independently deployable stages and whose edges encode explicit data dependencies.The graph is G = (V, U), with each vertex v ∈ V representing a stage and each edge u ∈ U representing a dependency.
Unified stage-level abstraction: Every stage must be serializable, relocatable, and independently executable to support device-edge deployment.These conditions require structured intermediate packets, pre-loaded or resident execution context, and no implicit cross-stage state.
Unified stage-level abstraction: The abstraction represents autoregressive and diffusion-based VLAs as attributed stage-level DAGs despite their different execution states.Autoregressive models carry sequential context and KV-cache state, whereas diffusion-based models propagate latent representations and iterative denoising state.
Co-inference mapping space: EcoVLA formalizes co-inference plans as mappings Π = {π | π : V →D} from inference stages to available devices.The representative deployment scenario uses multiple heterogeneous robot devices collaborating with a single shared edge server.
Co-inference mapping space: The scheduler evaluates only deployable mappings, restricting candidates to stages executable on assigned devices based on memory, operator support, and data-type compatibility.Feasible plans are confined to the deployable subspace through the executability condition C(v, d).

5 System Performance Awareness

EcoVLA builds a unified performance model that combines offline stage profiling, online communication and workload sensing, and post-execution feedback to estimate candidate mappings without exhaustive trial-and-error. It jointly predicts latency, energy, and energy efficiency while accounting for computation, transmission, queuing, and runtime conditions.

Model construction: EcoVLA combines offline stage-level profiling, online communication and workload sensing, and post-execution feedback to continuously calibrate candidate-mapping performance estimates.This avoids the prohibitive overhead of evaluating every mapping online in heterogeneous device-edge systems.
Stage-level modeling: For each stage-device pair, predicted execution time sums steady-state computation and boundary overhead, while stage energy equals effective power multiplied by predicted execution time.Boundary overhead includes packet encoding and decoding, dispatch, queuing, and send-side processing.
Communication and workload awareness: EcoVLA estimates cross-device transfer cost from each structured intermediate packet’s actual byte size and the current data-plane probing cache.Edge workload prediction additionally uses request inter-arrival times, SLO constraints, and current queue states to estimate queuing delay.
Request-level estimation: EcoVLA combines stage profiling, communication awareness, and workload awareness into unified end-to-end latency and energy estimates for candidate mappings.Runtime load indicators include device availability and stage-associated queue states, while online residual feedback compensates for systematic modeling bias.
Energy-efficiency objective: Because action-chunk length is usually predetermined, energy-efficiency differences among candidate mappings are dominated by end-to-end energy cost.Lact(π) denotes the number of effective control steps produced by one inference request under mapping π.

6 Adaptive Scheduling

EcoVLA’s adaptive scheduler selects the most energy-efficient feasible co-inference plan for each request while meeting a configurable real-time latency budget. It evaluates candidate execution plans using predicted latency, energy efficiency, transmission costs, node availability, and queue states, then reserves the chosen plan’s occupancy.

Scheduling objective: EcoVLA defines scheduling as maximizing predicted energy efficiency among candidate plans whose predicted latency meets the per-request SLO budget TSLO.TSLO is determined by target control frequency and action chunk length, and users can adjust it to trade stricter real-time guarantees against higher system energy efficiency.
Candidate evaluation: For each request, EcoVLA enumerates plans respecting device capability, stage residency, and cold-start constraints, then recursively estimates stage timing from latency, energy, transmission, availability, and queue-state profiles.For batching-capable edge stages, the scheduler also estimates batching-related execution behavior.
Plan selection: After evaluation, the scheduler filters candidates satisfying ˆT(π) ≤ TSLO, selects the highest predicted-energy-efficiency plan, and otherwise falls back to the candidate with the smallest predicted latency.If no candidate satisfies the SLO under the current system state, the scheduler minimizes the violation magnitude.
Runtime integration: Once selected, the scheduler registers node-occupancy intervals as reservations so subsequent requests perceive short-term committed load.The runtime preserves stage boundaries, queue semantics, and batching opportunities to keep scheduling decisions valid during execution.
Runtime integration: EcoVLA implements schedule-preserving stage execution and communication middleware that separates control metadata from inter-stage tensor payloads across independent channels.Same-node edges use local inline passing, while cross-node edges transmit each packet’s tensor tree.

7 Experimental Evaluation

EcoVLA is evaluated on five VLA models using a real device-edge testbed under a 20 Hz action-output constraint. It achieves higher energy efficiency than feasible baselines and adapts to network and workload variability through runtime scheduling.

Models, baselines, and request setup: Five representative VLA models are evaluated, including OpenVLA, π0, π0.5, SmolVLA, and RDT-1b.Diffusion-based models use an action chunk size of 20, while single-robot baselines use official inference interfaces.
Device-edge system and implementation setup: The real device-edge testbed pairs an NVIDIA RTX 4090 edge server with an NVIDIA Jetson AGX Orin 32GB robot device.EcoVLA uses Ray RPC for control, torch.distributed with Gloo for data, preloads stage weights, and profiles power on both nodes.
Real-time constraint: 20 Hz defines the real-time SLO, corresponding to per-inference latency TSLO ≤1 s for action chunks of 20.This control frequency follows prior work.
EcoVLA vs. Existing approaches: 236% improvement over Fixed Split is achieved by EcoVLA on OpenVLA, with gains of 58%, 84%, 100%, and 182% on π0, π0.5, SmolVLA, and RDT-1b.EcoVLA consistently achieves higher energy efficiency among SLO-satisfying schemes.
Evaluation under Network and Edge Workload Variability: 0.1 s latency and 2.4 Actions/J are maintained during degraded bandwidth after EcoVLA detects the change and reschedules to a device-heavier plan.Bandwidth is restricted from 200Mbps to 50Mbps; Fixed Split degrades most severely because its static communication budget no longer holds.
Evaluation of System Performance Awareness: EcoVLA’s system-performance-awareness module is evaluated by comparing predicted and measured per-request latency and energy across candidate mappings.The evaluation also reports prediction accuracy under different error-tolerance bounds.

8 Conclusion

EcoVLA is presented as a paradigm-agnostic adaptive device-edge co-inference framework for maximizing VLA energy efficiency under real-time constraints. It unifies heterogeneous VLA computation graphs through stage-level abstraction, while future work targets multi-edge collaboration and VLA-specific compression.

EcoVLA is the first paradigm-agnostic adaptive device-edge co-inference framework for VLA models targeting energy-efficiency maximization under real-time constraints.
EcoVLA decouples heterogeneous VLA computation graphs into a unified stage-level abstraction, establishing an architecture-agnostic co-inference design space and structured intermediate-state interface.
Future work will extend EcoVLA to multi-edge collaborative scenarios and incorporate VLA-specific compression techniques into the stage-level co-inference design space.These extensions aim to further push the energy-efficiency frontier of real-time embodied AI deployment.