Latency-Tolerant Cloud-Edge Collaborative Vision-Language-Action Models via Emergent Representational Specialization

Daojie Peng; Fulong Ma; Bingtao Wang; Sheng Wang; Jun Ma

AAAI 2027 · Under Review

CloudEdgeVLA

Cloud reasoning. Edge reflexes. No waiting.

Latency-Tolerant Cloud-Edge Collaborative Vision-Language-Action Models via Emergent Representational Specialization

Daojie Peng¹ Fulong Ma¹ Bingtao Wang² Sheng Wang³ Jun Ma^1,†

¹HKUST(GZ) ²Shandong University ³RoboScience ^†Corresponding author

arXiv Explore the method

90.8% normalized delay AUC

76.5% success retained under a 40-step window

Abstract

Latency tolerance as a learned interface.

Deploying billion-parameter VLA policies on mobile robots creates a systems conflict: semantic reasoning benefits from cloud GPUs, while closed-loop control must respond locally despite delay and jitter.

CloudEdgeVLA separates information by temporal role. A cloud VLA turns delayed observations into slowly varying task features, while a lightweight edge head grounds the latest available feature with current local vision. Paired-frame dual-path training teaches both fresh and delayed cloud features to predict the current action, encouraging temporal specialization without requiring delay metadata at inference.

Non-blocking control

The edge always acts with the newest feature it has; it never waits for a particular cloud response.

Paired-frame training

Fresh and randomly delayed cloud frames share the same current action target.

Emergent specialization

The cloud preserves task context while the edge supplies state-sensitive corrections.

The deployment gap

A fast robot cannot wait for a slow round trip.

Action chunks grow stale, and synchronized fast-slow systems still assume bounded timing. CloudEdgeVLA makes cloud updates independent from the edge control clock.

Comparison of action chunking, synchronized fast-slow control, and asynchronous CloudEdgeVLA — **Three dual-system paradigms.** CloudEdgeVLA is the only design shown here whose edge execution is independent of the cloud update clock.

Architecture

Slow semantics meet current state.

A single learned interface separates latency-insensitive planning from latency-sensitive control, allowing both sides to scale on their own cadence.

CloudEdgeVLA asynchronous deployment architecture and paired-frame dual-path training pipeline — **Deployment and training.** The edge never blocks on a cloud response. During training, the fresh and stale cloud paths share the same current edge feature z_t and current action target a_t.

Cloud · System 2

Encode what to do

The 7B VLA backbone processes a delayed observation and language instruction into reusable planning features.

h_t-k = f_θ(o_t-k, ℓ)

Edge · System 1

See what is true now

A frozen lightweight SigLIP encoder observes the current frame locally, with no network delay in its signal.

z_t = v_ψ(o_t)

Action · Closed loop

Ground and act

The edge action head fuses old task context with current visual evidence and executes immediately.

â_t = g_φ(h_t-k, z_t)

Paired-frame dual-path training

Teach the feature to survive age.

For each current frame, training samples a delayed frame from the same episode. Both cloud features are fused with the same current edge vision and supervised toward the same current action.

ℒ = (1 − λ) ‖â_fresh − a_t‖₁ + λ ‖â_stale − a_t‖₁

No explicit invariance loss. No delay value is required at inference.

Fresh path Current cloud frame o_t → h_fresh current edge vision z_t â_fresh → a_t

Stale path Randomly delayed frame o_t-d → h_stale current edge vision z_t â_stale → a_t

LIBERO results

The performance gap opens with delay.

Explore success rates across four LIBERO suites. The d_max = 0 point is the synchronous reference; each non-zero setting samples observation age uniformly from 1 to d_max. CloudEdgeVLA values average three seeds.

Uniform delay window d_max = 40 steps

CloudEdgeVLA

73.5%

VLASH

1.7%

UniVLA

0.9%

OpenVLA-OFT

0.0%

OpenVLA

0.0%

+71.8 points over the best baseline at 40 steps, averaged across all suites.

Beyond the training window

90.8% Normalized delay AUC VLASH: 32.4%

73.5% Mean success at 40 steps Best baseline: 1.7%

63.8–78.0% Suite-level success at 40 steps Across Spatial, Object, Goal, and Long

Delay sweep, normalized delay AUC, and retention score for CloudEdgeVLA and baseline methods — **Closed-loop delay sweep.** CloudEdgeVLA remains strong beyond the 20-step training window and averages 73.5% success at 40 steps.

View suite-level results

LIBERO task success (%) under uniformly sampled observation delay.
Method	Delay	Spatial	Object	Goal	Long
OpenVLA	0	84.6	71.2	77.0	56.2
	10	6.8	1.2	2.2	5.2
	20	0.2	0.0	0.0	0.0
	40	0.0	0.0	0.0	0.0
OpenVLA-OFT	0	98.4	98.6	97.2	93.4
	10	10.6	3.6	26.2	8.4
	20	0.0	0.0	4.0	1.4
	40	0.0	0.0	0.0	0.0
UniVLA	0	96.0	96.6	94.6	93.2
	10	27.4	34.8	48.2	35.0
	20	0.4	2.6	21.8	2.2
	40	0.0	0.0	3.0	0.5
VLASH	0	97.3	99.6	96.7	93.5
	10	60.0	62.8	56.8	45.6
	20	5.0	8.2	28.2	9.2
	40	0.0	0.4	6.4	0.0
CloudEdgeVLA	0	97.9	97.8	96.5	91.7
	10	93.6	94.4	92.9	83.2
	20	89.6	92.1	90.9	78.1
	40	76.4	75.6	78.0	63.8

CloudEdgeVLA values are means over seeds 7, 8, and 9. Standard deviations are reported in the paper.

Why it works

Robustness appears on both sides of the interface.

Under a 20-step delay, the learned cloud representation changes less, and the edge head transfers less residual feature drift into the executed action.

59.2%: less backbone drift
72.7%: less staleness transfer

Backbone hidden-state cosine distance under increasing image delay — **Backbone staleness.** At 20 steps, CloudEdgeVLA reduces hidden-state drift by 59.2%.

Action head staleness-transfer gain under increasing image delay — **Head transfer.** At 20 steps, the action head transfers 72.7% less residual feature drift into actions.

Real-world deployment

The edge loop keeps moving with 1,000 ms added RTT.

A Franka robot places a toy bear into a box in static and dynamic-target settings. The cloud 7B backbone runs on an RTX 4090; edge vision, action head, and control run on an RTX 5080 workstation.

Simulation comparison and CloudEdgeVLA real-robot rollout sequence — Simulation comparison above; static and dynamically displaced real-robot targets below.

Static / dynamic success over 10 trials
Added RTT	VLASH	CloudEdgeVLA
0 ms	100 / 90	100 / 90
400 ms	60 / 30	90 / 90
1,000 ms	0 / 0	80 / 70

Each cell reports static / dynamic success over 10 trials per variant; added RTT is emulated on top of the native serving pipeline.

Citation

Build on CloudEdgeVLA.

Please cite the paper if this work supports your research.

Read on arXiv Manuscript PDF

@misc{peng2026latencytolerantcloudedgecollaborativevisionlanguageaction,
  title         = {Latency-Tolerant Cloud-Edge Collaborative
                   Vision-Language-Action Models via Emergent
                   Representational Specialization},
  author        = {Daojie Peng and Fulong Ma and Bingtao Wang
                   and Sheng Wang and Jun Ma},
  year          = {2026},
  eprint        = {2608.00569},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.00569}
}