Compute Regeneration through Re-Productization: Reorganizing Retired Edge AI Silicon into a Delivered Multilingual Service Robot

Hongyu Luo; Sun Tie; John Wang

cs.RONov 8, 2026

Compute Regeneration through Re-Productization: Reorganizing Retired Edge AI Silicon into a Delivered Multilingual Service Robot

Authors: Hongyu Luo, Sun Tie, John Wang

Organizations: Beijing Data Chain Technology Co., Ltd. Beijing, China · University of Science and Technology Beijing Beijing, China

Abstract

Useful edge hardware can lose product support while retaining substantial computation. We treat this as systems design: compute regeneration here means re-productization, retaining the compute module while redesigning the surrounding product organization. Our case is GEdgeBot, a multilingual service robot built around a retired 32 GB Jetson AGX Xavier. A source-grounded inventory shows the box alone supplies inference, while the combined system exposes five on-device AI services, a seven-language speech chain, three sensing domains, and one deterministic seven-byte control boundary. AI and orchestration stay on the Xavier; finite motor-pulse generation on an STM32F407. Audits cover the six-mode control state machine, readiness barriers and sound orientation. Historical single-prototype measurements report 1,782 MB GPU memory with YOLO, Whisper-small, and NLLB-200 resident, a direct-control p99 deviation rising from 0.24 to 49.93 ms under concurrent AI load, and six of six trigger-to-follow demonstrations at 4.35 s mean latency, with conversational voice unavailable. A November campaign restores an isolated inference stack, re-measures AI workloads, and repeats control-frame runs. Driving two execution organisations with the same 6,000 measured 50 ms frame intervals shows the finite-pulse boundary removes host jitter from the pulse train but not the command start phase; its stopping bound is host-enforced, while the MCU 16-bit pulse field is unclamped. A six-check static audit exposes a host--MCU queue mismatch, and a 14-scenario replay executes the audited main() dispatcher across six modes with zero physical device opens and 54 contract-checked frames. Results support an existence claim and a reproducible architectural hypothesis, not fleet-level reliability. The 208 direct-control ratio remains unreplicated; MCU waveform capture and on-site fault injection remain open.

Figures & tables

Panel of Figure 1 — Fig. 1: The proposed value-capture ladder is a systems-design distinction, not a measured ranking of economic or environmental benefit. Re-productization preserves the compute module while redesigning the product organization around it; in this case an AI box becomes the compute core of a delivered service robot.

Product function	Box alone	Box + robot: what the integration adds
On-device inference (YOLO / Whisper / NLLB / ChatTTS)	yes †	retained as the compute core; isolated CUDA stack on the frozen BSP
Seven-language ASR $\to$ MT $\to$ TTS	models only	five on-device services; microphone array, prewarm/readiness, orchestration
Sound orientation and counter-turn	no	recovered DOA protocol and seven-byte UART frames
Two-level person following	no	OpenNI camera and head/body pulse state machine
FPD-Link camera registration	no	two kernel driver modules and a V4L2 graph
Finite-pulse bounded actuation	no	STM32F407 and the seven-byte frame

TABLE I: Product functions available from the retired compute module alone versus the module embedded in the delivered robot. Source-grounded integration inventory, derived from the captured source tree and the November device JSONs.

Panel of Figure 3 — Fig. 2: Source-grounded GEdgeBot architecture. The Astra/OpenNI path is used by the captured person-following runtime; the FPD-Link path is a separately documented V4L2 integration and is not assumed to be the camera used in every demonstration. Online connectors are shown explicitly because the current voice assistant is not invariably offline.

Control element	Derived count	What the combination adds
CLI modes / dispatch	6 / 6	six operating modes around one retained accelerator
Concurrent contexts	4 threads, 2 subprocesses, 1 HTTP server, 7 HTTP clients	voice process, camera capture, monitor, UI/API sidecars
Readiness barrier	2 gates $\times$ 60 s	prewarm hides multi-second model cold starts before wake-up
Degraded branches	97 branch sites, 20 categories	readiness failure, API-vision warning, missing angle, dead zone, dropout, dry-run
Safety interlocks	6 methods	dead zone, two-frame confirmation, rate limits, settle, dry-run default
Control boundary	7 bytes, 3 opcodes	one narrow semantic interface to the STM32F407

TABLE II: Overall control plane of the audited runtime (2026-11-07). AST-derived inventory of the captured source snapshot; branch presence is not branch execution, and no physical measurement is claimed.

Panel of Figure 5 — Fig. 3: Overall control state machine of the audited RobotControl/robot_runtime.py auto-wakeup mode. The prewarm and readiness barrier sit outside the trigger-to-follow path; the box-plus-robot integration is what introduces the concurrent voice process, camera thread, HTTP monitor, microphone array and seven-byte actuator boundary that this state machine orchestrates. Source-derived structure, not a runtime trace.

Panel of Figure 6 — Fig. 4: Host-side intent of the sound-orientation callback in orient_to_sound . The drawing expresses intended kinematic cancellation, not a measured trajectory. The audited MCU’s blocking receive/reset logic does not reliably queue the second host command; its in-routine counter-turn is fixed at 15 pulses.

Workload	Reported result
YOLOv8n@320, 15 W	26.6 FPS; mean 37.6 ms; p99 84.0 ms
YOLOv8n@640, 15 W	16.3 FPS; mean 61.2 ms
Whisper-small, 10 s audio	RTF 0.091; 2 s clip RTF 0.44
NLLB-200, short/mid/long	0.42/1.02/1.55 s in the single-model table
ChatTTS, long text	RTF 2.51; not real time
Three models resident	1,782 MB GPU memory

TABLE III: Historical single-prototype AI results (October 2026). Values are transcribed, not independently re-derived in this draft.

Metric	MAXN	15 W
YOLOv8n@320 (ms)	$23.40\pm 6.23$ †	$18.65\pm 0.03$
YOLOv8n@640 (ms)	$22.14\pm 1.24$	$24.44\pm 0.26$
Whisper 10 s RTF	$0.159\pm 0.003$	$0.215\pm 0.008$
NLLB short, beam 5 (s)	$0.735\pm 0.010$	$0.762\pm 0.019$
NLLB short, greedy (s)	$0.463\pm 0.004$	$0.480\pm 0.015$
ChatTTS long-text RTF	2.982	—

TABLE IV: New November 2026 measurements on the same Xavier, using audited source parameters (

n=3

repeats per power mode). An independent batch: not a reproduction of the October values and never pooled with Table III .

Condition	Std (ms)	p99 dev. (ms)	Max (ms)
Oct., idle, 15 W †	0.063	0.24	—
Oct., AI, 15 W †	13.37	49.93	212 †
Oct., AI, MAXN †	11.41	49.92	182 †
Nov., idle, MAXN, pty	0.216	0.246	52.136

TABLE V: Control-timing records. The November run is a different, no-motion protocol and must not be pooled with October data.

Transport	Load	$n$	p99 range (ms)
Real /dev/ttyUSB0	Idle	3	0.093–0.480
Real /dev/ttyUSB0	YOLO@640	3	0.095–0.104
pty (Section V)	Idle	5	0.051–0.871
pty (Section V)	YOLO@640	5	0.064–2.908

TABLE VI: Command start-phase deviation on the real motor UART and on the pty transport, all November 2026 no-motion measurements with the same estimator. Real-port frames are inert; the historical pulse-train statistic is not reproduced here.

Condition	$n$ / completed	Orient budget (s)
Nominal	6 / 6	$4.211\pm 0.144$
Voice degraded	6 / 6	$4.211\pm 0.144$
Detector dropout	6 / 6	$4.211\pm 0.144$
DOA dead zone	6 / 6	0
Camera not ready	6 / 0	$4.211\pm 0.144$

TABLE VII: Software-in-the-loop product-loop replay (2026-11-07): 30 traces driven by synthetic events and a virtual clock. Not a physical demonstration and not a reproduction of the October runs.

Measure	Result	Interpretation
Scenarios / modes	14/14, 6/6	every audited entry point executes
Auto-wakeup cycles	10	prewarm $\to$ trigger $\to$ orient $\to$ voice $\to$ follow $\to$ flush
Frames / errors	54 / 0	seven-byte contract and host clamps hold
Physical / logical opens	0 / 33	guards fired 0 times; stubs recorded calls
State transitions	48 distinct	ordered, machine-readable control trace
Degraded branches	8 exercised	readiness, API warning, dead zone, dropout, dry-run, no angle, skip-head

TABLE VIII: Full-runtime control-state replay (2026-11-07): 14 offline scenarios executing the real six-mode dispatcher. Not a physical demonstration and not a reproduction of the October runs.

Item	Design/historical text	Captured source and implication
Motor integrity	Abstract: sequence and CRC	Seven bytes; no sequence/CRC; header/tail only
Turn opcodes	Patent: left 05 , right 06	Runtime: left 01 , right 02
Head–body delay	Patent: 0.4 s	Runtime default: 2.0 s
Sound body limit	Patent example: up to 60 pulses	Runtime default maximum: 25 pulses
Small follow pulse	Patent: 3 pulses	Runtime selects 5 and enforces a minimum of 5
Counter-turn path	Patent: equal planned opposite pulse count	MCU uses fixed 15 in routine; buffered host frame can be discarded

TABLE IX: Design-document and captured-source differences. Different entries may describe different revisions; none is silently treated as the October execution record.

Real-Time Robot Control

See all 32 papers

May 21, 2026

$E^3$ -Agent: An Executable and Evolving Agent for Resource Management of Edge Generative Inference

Edge deployments of generative inference increasingly face two practical realities: per-device per-model performance is often unknown at deployment time, and it is non-stationary due to user-driven semantic events, background load, and device churn. Consequently, a resource manager that is tuned offline under a fixed regime can become brittle and expensive to maintain. This paper presents

E^3

-Agent, an executable and evolving agent for edge artificial intelligence generated content (AIGC) resource management.

E^3

-Agent separates a fast-path router that makes millisecond-level dispatch decisions from a slow-path, event-driven large language model (LLM) meta-controller that mitigates regime shifts through a small, explicit control surface exposed via a tool interface, including risk gating, router configuration, and rapid performance calibration. The agent learns online from execution feedback and continuously adapts to unknown and time-varying service-time mappings. We evaluate

E^3

-Agent in a discrete-event simulator driven by MLPerf-derived device-model measurement priors, covering cold-start warmup and three dynamic regimes: semantic dynamics, device churn, and hidden drift. Across the dynamic scenarios,

E^3

-Agent reduces average latency by 65%-73% compared to the best static baseline, stays within 7%-10% of an online full-information Oracle used for evaluation, and effectively suppresses stutter rate under semantic degradation.

Rui Bao, Yaping Sun, Zhiyong Chen +4

Cooperative Medianet Innovation Center, Shanghai Jiao Tong University, Shanghai 200240, China · Department of Broadband Communication, Pengcheng Laboratory, Shenzhen 518000, China · China Mobile Research Institute, Beijing 100053, China

LLM Inference Efficiency Edge Inference Edge Computing LLM Agent Orchestration

Ask PDF

Sep 7, 2026

Beyond Fluent Generation: A CPU Reliability Benchmark for MCP-Style Tool Calling in Sub-2B Small Language Models for Edge Deployment

Resource constrained single-board computers including Raspberry Pi, NVIDIA Jetson Nano, Arduino UNO Q, Orange Pi, and LattePanda motivate on-device small language model (SLM) agents that reduce cloud dependence, improve data locality, and tolerate intermittent connectivity. Model Context Protocol (MCP)-style tool invocation demands more than fluent generation: an agent must emit machine-readable JSON, select the correct tool, supply all required arguments, and avoid unintended actions. We establish a platform-agnostic CPU baseline by evaluating five open-weight models below two billion parameters Phi-1.5, Pythia-1.4B, TinyLlama-1.1B-Chat, Qwen2.5-0.5B, and Qwen2.5-1.5B on 100 prompts spanning weather retrieval, web search, calculation, email composition, and task creation, under greedy decoding and nucleus sampling. A recovery parser strips Markdown fences, extracts brace-delimited substrings, and scores parseability, tool-name correctness, argument completeness, and value agreement. Under this criterion, Qwen2.5-1.5B achieves 75% (greedy) and 79% (sampling); Qwen2.5-0.5B achieves 72% (greedy) but drops to 32% under sampling. Phi-1.5 scores 0%; Pythia and TinyLlama reach at most 7%. A strict post-hoc audit finds only 5 of 1,000 raw responses directly parseable as JSON, exposing near-total dependence on output recovery. A CPU resource probe shows Qwen2.5-1.5B requires 7,960 MiB and 30.782 s mean latency; Qwen2.5-0.5B uses 3,637 MiB and 10.627 s, revealing a reliability-resource trade-off for edge deployment. These results do not cover the named boards directly or a full MCP implementation. Safe deployment requires schema validation, constrained generation, least-privilege execution, and human escalation for consequential actions.

Abrar Shahriar Qurat-Ul-Ain Mastoi

University of the West of England, Bristol,BS16 1QY United Kingdom

Tool-Augmented Language Model Agents On-Device Language Model Inference Language Model Generation Evaluation Small Language Models LLM Reliability LLM Tool Use+1

Ask PDF

Mar 1, 2026

Characterizing VLA Models: Identifying the Action Generation Bottleneck for Edge AI Architectures

Vision-Language-Action (VLA) models are an emerging class of workloads critical for robotics and embodied AI at the edge. As these models scale, they demonstrate significant capability gains, yet they must be deployed locally to meet the strict latency requirements of real-time applications. This paper characterizes VLA performance on two generations of edge hardware, viz. the Nvidia Jetson Orin and Thor platforms. Using MolmoAct-7B, a state-of-the-art VLA model, we identify a primary execution bottleneck: up to 75% of end-to-end latency is consumed by the memory-bound action-generation phase. Through analytical modeling and simulations, we project the hardware requirements for scaling to 100B parameter models. We also explore the impact of high-bandwidth memory technologies and processing-in-memory (PIM) as promising future pathways in edge systems for embodied AI.

Manoj Vishwanathan, Suvinay Subramanian, Anand Raghunathan

Google, Mountain View, CA USA · Purdue University, West Lafayette, IN USA

Vision-Language-Action Models Efficient VLA Model Inference AI Accelerator Inference Edge Computing

Ask PDF

Abstract

Real-Time Robot Control

Explore similar work

E3E^3E3-Agent: An Executable and Evolving Agent for Resource Management of Edge Generative Inference

Beyond Fluent Generation: A CPU Reliability Benchmark for MCP-Style Tool Calling in Sub-2B Small Language Models for Edge Deployment

Characterizing VLA Models: Identifying the Action Generation Bottleneck for Edge AI Architectures

$E^3$ -Agent: An Executable and Evolving Agent for Resource Management of Edge Generative Inference