Material Synthesis AI Prediction and Multi-Agent Embodied Lab Assistant

Dual RDK X5 embedded robot system for the D-Robotics embedded contest.

Overview

This project builds an embedded laboratory assistant for near-infrared phosphor material research. The system connects material formula screening, XRD/PL evidence analysis, embodied execution, workstation manipulation, and public evidence display into one traceable engineering loop.

The submitted public repository is a reviewable boundary of the real project. It shows the system design, frontend evidence platform, interface schema, report assets, non-core workstation logic, edge-facing adapters, public screenshots, public crystal cache files, and offline demo sensor frames. It does not publish private data, model weights, credentials, deployment scripts, or the core material prediction engine.

System Design

The complete system is organized as four cooperating layers:

Layer	Main hardware	Responsibility
AI brain	RDK X5, 4K camera, audio module	Material prediction, XRD/PL analysis, local LLM reasoning, evidence logging
Embodied brain	RDK X5, LD14 LiDAR, Astra depth camera, odometry	Verified pickup/lift/0.50 m odometry loop/release/reset, SLAM, and Lab-FSD shadow/assist
Workstation	Dual myCobot 280-Pi stations, independent cameras, custom gripper	arm01 visual redundancy and bag drop, arm02 concurrent four-cycle grinding
Execution layer	STM32F407, servo, electric push rod, electromagnet, stepper axis	Low-level timing, bottle pickup/release, safety-degraded actuator sequence

As of July 20, 2026, the first three finals segments have been rehearsed on hardware and frozen. The embodied brain completed bottle pickup, lift, a 0.50 m odometry-closed straight drive, bottle release, and reset. The existing AI Dashboard, XRD vision, and material-synthesis prediction flow completed a tablet rehearsal. The dual-arm station completed arm01 visual redundancy and bag drop together with arm02 concurrent four-cycle grinding. Lab-FSD remains shadow/assist and has no chassis authority. X5 CPU/OpenCV is authoritative for bag state; BPU output is assist/evidence only. Neither the public portal nor this repository exposes hardware actuation.

Key Capabilities

• Dual RDK X5 heterogeneous cooperation: one AI brain for research reasoning, one embodied brain for mobile perception and planning.
• Four AI analysis lines: vision, XRD numerical analysis, PL vision, and PL numerical analysis.
• Local embedded inference boundary: BPU lightweight models, CPU local LLM processes, and offline fallback paths.
• Verified embodied hardware loop: pickup, lift, 0.50 m odometry-closed drive, release, and reset.
• Lab-FSD shadow planner: FSD-style BEV occupancy reasoning for risk, trajectory, and safety-gate output without direct chassis takeover.
• Finals dual-arm collaboration: arm01 visual redundancy and bag drop with arm02 concurrent four-cycle grinding; CPU/OpenCV is authoritative and BPU remains assist-only.
• STM32F407 actuator sequence: servo, electric push rod, electromagnet, and stepper axis timing under safety-degraded demonstration rules.
• Public evidence platform: static frontend, OpenAPI-style schemas, report figures, rendered pages, screenshots, and reproducible review assets.

Repository Map

Path	Purpose
public_site_static/	Static frontend of the public evidence site, including PWA and 3D crystal display assets
public_site_reports/	Read-only public status report samples
public_site_tools/	Public 3D crystal asset generation scripts
workstation_public/	Non-core workstation interlock, mock telemetry, skill replay, and icon tooling
workstation_frontend_public/	Public workstation UI components, charts, 3D arm scenes, stores, and pages
edge_public/	Public interface stubs for Lab-FSD shadow planning, Fly-MB decision output, and F407 timing
schemas/	Read-only status API schema and example response
report_source/	TeX report source, HTML figure source, MATLAB/Python figure scripts, generated figures
public_evidence_data/	Public screenshots, rendered report pages, public crystal cache files, and offline demo sensor frames
docs/	Public project map and open boundary notes

Public Evidence

The repository includes evidence assets that can be inspected without private credentials:

• 23 rendered report pages in public_evidence_data/report_rendered_pages/.
• Public evidence site screenshots in public_evidence_data/site_screenshots/.
• Public crystal structure cache files in public_evidence_data/crystal_public_cache/.
• Offline replay-style sensor frames in public_evidence_data/demo_sensor_frames/.
• Static report figures generated from HTML, MATLAB, and Python sources.

How To Inspect

This public repository is designed for review rather than direct robot deployment.

1. Open public_site_static/index.html for the archived offline evidence-site snapshot; the current read-only portal is https://xiaomiju.xyz.
2. Read report_source/main.tex and report_source/sections/ for the complete design report source.
3. Inspect schemas/openapi_status_schema.json and schemas/status_snapshot_example.json for the public status API shape.
4. Review workstation_public/ and edge_public/ for non-core logic and interface boundaries.
5. Use the rendered pages and screenshots under public_evidence_data/ as offline evidence.

NodeHub Information Draft

• Project name: Material Synthesis AI Prediction and Multi-Agent Embodied Lab Assistant
• Chinese project name: 基于双 RDK X5 异构协同的材料合成 AI 预测与多机具身实验助理机器人
• Repository: https://github.com/Xiaomiju-x/xrd
• Suggested tags: RDK X5, Embedded AI, Materials AI, Laboratory Robot, SLAM, BPU, STM32F407
• Suggested platform: RDK X5
• Suggested category: Robot application, Embodied AI, AI vision, Research automation
• Video: submit the Bilibili embed code after the contest video is published.

Open Boundary And License

The public repository uses an Apache-2.0 open-source boundary for non-core review materials. Private laboratory assets remain excluded:

• API keys, cookies, SSH/WiFi/SSO configuration, account credentials, and private network addresses.
• GGUF, LoRA, BPU bin, tokenizer, tensor packages, and other model weights.
• Unpublished XRD/PL raw data, private experiment logs, failure-pattern libraries, and complete training sets.
• Core material prediction engine, private rules, deployment scripts, and executable real-hardware control scripts.

See docs/PUBLIC_BOUNDARY.md for the detailed publication boundary.