Important
Grand Prize: $8,000 USD Cash + Global UiPath Recognition! Let's automate the enterprise with fütüristik AI Orchestration! 🤖💼
An open-source framework for running a cloud-GPU orchestrator and humanoid RL training lifecycle manager as an agentic, human-in-the-loop (HITL) orchestration on UiPath Maestro BPMN — combining UiPath's two signature strengths: agentic AI (Agent Builder decisioning) and RPA (robots that provision cloud GPUs and execute into back-office systems).
Training neural skills (walking, running, grasping) for bipedal humanoid robots (like the Unitree G1) in reinforcement-learning simulation is compute-heavy and expensive. Teams burn thousands of dollars on cloud GPUs (AWS, GCP, AMD Developer Cloud) that sit idle, or on RL runs that silently diverge or collapse for hours before anyone notices. This project orchestrates the whole training lifecycle — from a researcher's training request, to cost-optimal GPU provisioning, to live anomaly detection that kills wasted runs, to human-approved checkpoint registration — as a BPMN process where AI agents do the analysis and humans approve the critical cost and physics-safety gates.
| Stage | Project | Role |
|---|---|---|
| Deploy | gcp |
Pick the right robot for a task + plan the install |
| Operate | gemini-xprize — FleetMind |
Run the autonomous day-to-day ops loop |
| Govern | uipath-maestro (this repo) |
Manage training runs + cloud GPU budget with human-approved gates (the governance / HITL layer) |
| Respond | protocol-sift-dfir |
Forensics when a unit is compromised |
Reference workload across all four: Unitree G1-class humanoid RL skills trained in MuJoCo.
- Agentic — the decisioning runs as UiPath Agent Builder agents: a Telemetry Analyst (classify the anomaly + Anomaly Risk score), a Cost & Impact Agent (cloud-GPU economic + organizational impact → routing), and a compliance-summary agent. They reason over each training-run anomaly and decide fast-track vs human review.
- RPA — once an action is approved, a UiPath RPA robot executes it into systems that have no clean API: cloud-provider consoles, the billing/quota dashboards, the model registry, and SSH-driven container control — driven through the UI when needed, plus API Workflows where an API exists.
- 8 process containers modeling the full training lifecycle (training request → GPU provisioning → telemetry/anomaly analysis → risk/cost routing → approved remediation & audit) as Maestro BPMN subprocesses.
- 9 roles as BPMN swimlanes — some realized as autonomous LLM agents (Telemetry Analyst, Anomaly Risk Agent), others as human-in-the-loop review boards via UiPath Action Center.
- An agentic "Fast Track" gateway — the Anomaly Risk Agent scores each event and bypasses the full review board for low-risk, contained anomalies (auto-remediate, e.g. release an idle instance).
- Vendor-agnostic sources — live reward/loss telemetry, container logs, a hardware monitor (GPU/thermal/quota), or a researcher console all normalize to one canonical
IncidentReport(training-run anomaly event). - Audit-grade traceability — immutable log entries at every gateway and task node (GPU hours, cost, approvals).
Process 1 · Anomaly & Request Intake
1. Anomaly Report Telemetry/logs/monitor → Telemetry Analyst → Researcher HITL → disposition
2. Manual Intervention Researcher console request or a fixed monitor flags an anomaly + feedback loop
Process 2 · Analysis & Risk Routing
3. Risk & Anomaly Analysis Anomaly → Anomaly Risk Agent → Fast Track OR Researcher Review Board (HITL)
4. Cost & Impact Analysis Cloud-GPU economic + organizational impact (idle spend / run loss / liability / fleet reach)
Process 3 · Action & Compliance
5. Remediation Plan Analyst Agent + RL Engineer draft a recovery playbook (retune/restart) → Ops HITL
6. Execute Remediation SSH-shutdown a diverging run, release/teardown an instance; HOLD a production-bound run
7. Cloud/Registry Write-back **RPA robot** writes to the cloud console, billing dashboard, model registry
8. Quality Assurance Compliance-summary agent + immutable audit log + researcher sign-off
Our architecture integrates the full suite of advanced UiPath Automation Cloud components as an enterprise governance and orchestration layer:
| UiPath Component | Description & Role in Solution |
|---|---|
| UiPath Maestro BPMN | The primary process orchestration engine. Models and controls the full 8-container training lifecycle from ingestion to termination/registry. |
| UiPath Agent Builder | Coordinates the specialized autonomous AI agents (Telemetry Analyst, Anomaly Risk Agent, Cost & Impact Agent) with structured prompt templates. |
| UiPath Action Center | Renders rich interactive Human-in-the-Loop (HITL) forms for researchers to inspect training telemetry (loss/reward curves) and approve/reject weight deployment. |
| UiPath RPA Robots & API Workflows | Executes hard integration tasks (such as GCP/AWS cloud-provider console interactions, SSH container termination, and model weight registry updates). |
| UiPath Orchestrator Queues | Manages the unified, high-reliability IncidentReports queue that triggers process executions upon new anomaly detection. |
This solution leverages a Hybrid (Both Coded and Low-code) agent architecture to combine enterprise-grade stability with high-intelligence reasoning:
- Low-code / No-code Agents (UiPath Maestro BPMN & Orchestrator): Handles the end-to-end routing, human approvals (Action Center), queue state transitions, and high-level execution policies.
- Coded Agents (Python / LangChain / Claude & Gemini): Handles real-time telemetry analysis (1Hz data processing, anomaly classification, gradient collapse detection, and risk scoring). These coded agents connect seamlessly to UiPath via REST API endpoints and Orchestrator Queues, creating a closed-loop system where advanced code-based intelligence runs under low-code BPMN governance.
Full architecture and design patterns: docs/ARCHITECTURE.md.
You can configure, run, and evaluate this solution using either a quick offline simulation or a full cloud deployment.
The entire agentic pipeline, telemetry monitoring, exclusive gateways, and write-back execution are fully testable locally with zero external dependencies (the LLM and cloud services fallback elegantly to rule-based mock logic if no API keys are provided):
-
Clone the repository and enter the directory:
git clone https://github.com/ForenlyAI/uipath-maestro-ecm.git cd uipath-maestro-ecm -
Initialize a virtual environment and install dependencies:
python3 -m venv .venv && source .venv/bin/activate pip install -r requirements.txt
-
Validate the JSON schemas and incident triggers: Verify that all incoming telemetry and resource alerts comply with the strict data model schema:
python -c " import json, glob from jsonschema import Draft202012Validator schema = json.load(open('samples/triggers/incident_report.schema.json')) v = Draft202012Validator(schema) for f in glob.glob('samples/triggers/*.json'): if 'schema' in f: continue v.validate(json.load(open(f))) print('Verified:', f) "
-
Run the complete end-to-end pipeline simulation: This runs the anomaly triage, executes the Coded Telemetry Analyst, routes the event via the Risk gateway, triggers the actions, and updates the local registry:
python run_pipeline.py
-
Run the automated test suite:
python -m pytest -v
To set up, configure, and execute the Maestro BPMN process live in the cloud environment:
- Studio Web Import: Connect your UiPath Studio Web workspace to this git repository, or import the pre-compiled BPMN diagram located at
maestro/FieldIncidentTriage/content/FieldIncidentTriage.bpmn. - Queue Setup: Create an Orchestrator Queue named
IncidentReportson your UiPath tenant and wire it as the process trigger. - HITL Forms & Integration: Deploy the Action Center schema (
artifacts/container1/action_center_irb.json) and configure your Agent Builder credentials. - Detailed Cloud Deployment Steps: For full step-by-step guidance on Automation Cloud provisioning, credential binding, and publishing, follow our comprehensive UiPath Cloud Deployment Runbook (docs/DEPLOY-MAESTRO.md).
Building this in the open — join the team chat on Discord: https://discord.gg/ntXbNbvN95
Contributions are welcome — see CONTRIBUTING.md. Day-to-day discussion happens on the project Discord.