### 1. Elevator Pitch (Slogan) — 194
Karakter (Limit < 200)
Cloud GPU orchestrator and humanoid
training manager on UiPath Maestro BPMN.
Automates ML on AWS/GCP/AMD, monitors
anomalies to prevent idle spend, and
routes checkpoints via HITL Action Center.
──────
### 2. Built With (Kullanılan
Teknolojiler Etiketleri)
uipath,uipath-maestro,uipath-agent-
builder,uipath-action-center,bpmn,
agentic-ai,cloud-gpu,humanoid-robotics,
reinforcement-learning,mujoco,aws,gcp,
amd-developer-cloud,python,docker,claude,
gemini
──────
### 3. Project Story (Proje Hikayesi) —
Markdown formatında
Aşağıdaki metni kopyalayıp Devpost'taki
Project Story alanına doğrudan
yapıştırabilirsiniz:
## Inspiration
Training advanced neural skills (such
as walking, running, or grasping) for bipedal humanoid robots (like the Unitree G1) in reinforcement learning (RL) simulation environments is incredibly compute-heavy and expensive. ML research teams waste thousands of dollars renting cloud GPU instances (AWS, GCP, AMD Developer Cloud) that run completely idle, or running RL training jobs that secretly diverge or collapse early without anyone noticing for hours.
Managing these asynchronous simulation
containers, monitoring live training telemetry, and protecting cloud budgets is a massive operational headache. UiPath AgentHack gave us the perfect opportunity to bring order to this chaos—building an orchestration backbone on UiPath Maestro autonomous, governed, and agentic BPMN to automate and optimize the entire
humanoid training lifecycle. ## What it does
**Training Fleet and Cloud Resource
Manager** is an autonomous cloud GPU orchestrator and humanoid RL training lifecycle manager built end-to-end on UiPath Maestro BPMN.
The real-world operational flow
operates as follows:
1. **Training Ingestion:** A robotics
run (e.g., walking) via a simple form,
researcher requests a new skill training
specifying target metrics.
2. Cloud Resource Check: An RPA
robot queries live APIs across AWS, GCP,
and AMD Developer Cloud to locate the
most cost-effective and available high-
performance GPUs (like AMD MI300X or
NVIDIA H100).
3. Container Deployment: Once
selected, the system spins up the
simulation container (g1-mujoco-rl-
training) on the remote cloud instance
via secure SSH.
4. Telemetry & AI Anomaly
Detection: The system continuously
monitors simulation logs and reward
curves. If an AI Quality Agent detects a
training anomaly (e.g., gradient collapse,
loss divergence, motor heat spikes, or
continuous falling loops), it triggers an
emergency SSH shutdown to instantly stop
wasting expensive GPU dollars and fires
an alert.
5. HITL Checkpoints & Action
Center: Successful checkpoints or
completed models generate an interactive
card in UiPath Action Center. The
researcher reviews the final reward
curves and joint heat maps to approve the
model, which triggers weight archival,
cloud-instance teardown, and resource
release.
6. Governance & Audit Logs: A
complete, auditable log of GPU
utilization, training efficiency, and
budget consumption is archived
automatically.
## How we built it
**Stack:**
- **UiPath Maestro BPMN** — the
orchestration core coordinating
simulation containers as subprocesses and
cloud providers as swimlanes.
- UiPath Agent Builder — powers the
Telemetry Analyst Agent (watching logs
and rewards for anomalies) and the
Budget/Risk Agent.
- UiPath Action Center —
interactive human-in-the-loop (HITL)
dashboards for researchers to review
reward curves and approve weight
registration.
- RPA Robots & API Workflows —
queries, and container SSH management
automates cloud provisioning, instance
across AWS, GCP, and AMD.
- Docker & MuJoCo / RL Environment
— the remote bipedal humanoid walking
training simulation containers (g1-
mujoco-rl-training).
- Claude (primary) + Gemini
(fallback) — LLM providers powering the
log analysis and reasoning behind
training health classification.
## Challenges we ran into
- **Standardizing Telemetry:**
Translating complex, high-frequency reinforcement learning simulation logs (MuJoCo/Stable-Baselines3) into a structured JSON schema that UiPath Maestro can seamlessly ingest and parse. - Low-Latency Terminations: Designing a highly reliable, instant SSH action pipeline to terminate cloud instances the moment training diverges, preventing runaway billing on expensive hardware. - Action Center Visualizations: Creating rich, lightweight data visualizations (reward curves, joint heat maps) that fit beautifully into UiPath Action Center for rapid researcher decision-making.
## Accomplishments that we're proud of
- **High-Performance Robotics meets
Enterprise BPMN:** Successfully bridging cutting-edge humanoid reinforcement learning pipelines with enterprise-grade process orchestration on UiPath Maestro BPMN. - Active Cost Containment: Realizing an autonomous "active cost containment" loop that can save robotics labs thousands of dollars in wasted cloud GPU hours by terminating bad runs instantly. - Generic, Anonymous Open Source: Building a fully generic, public, and open-source blueprint with no commercial claims or IP restrictions, making cloud GPU management accessible to any robotics lab.
## What we learned
- **BPMN for ML Pipelines:** BPMN 2.0
is exceptionally well-suited for long- running asynchronous ML jobs, which traditionally rely on fragile, custom- written shell scripts. - The True Power of HITL: Human-in- the-loop is vital in robotics, where AI- generated models must be verified against physics constraints before being pushed to physical hardware.
## What's next
- **Decentralized Compute
Integration:** Expand cloud integrations to include decentralized GPU networks (like Vast.ai, Akash) for even cheaper training. - Multi-Agent Co-Training: Support multi-agent reinforcement learning (MARL) where fleets of humanoid robots co-train in parallel simulation instances. - Multi-Framework Orchestration: Transition the prototype into an open- source, generic cloud orchestrator for any compute-heavy training workloads beyond robotics. ──────
Log in or sign up for Devpost to join the conversation.