Important
Prizes: State-of-the-art AMD Instinct / Radeon hardware & cash rewards to accelerate our AI models locally! Let's unlock raw performance! 🚀🔌
Entry for the AMD Developer Hackathon: ACT II (lablab.ai × AMD). "Build AI agents and high-performance AI apps on AMD GPUs in the cloud." Online · build Jul 6–11, 2026 · $10,000 prize pool.
GrootNet is a decentralized, distributed robotic brain network for autonomous robot skill acquisition, served across a fleet of AMD Instinct MI300X GPUs via ROCm.
Instead of deploying one monolithic model on one GPU, GrootNet splits the brain into specialized Groot 1.7 VLA (Vision-Language-Action) expert nodes — one per team member's cloud-allocated MI300X — and stitches them together at inference time into a single composite Frankenstein Model Node. Each member's expert node connects to this central Frankenstein node, which fans out every task, then fuses the experts' returned action logits into one unified command. A brain assembled from many distributed parts — hence Frankenstein.
Building this in the open — join the team chat on Discord: https://discord.gg/wWwUrHhyqg
Centered on AI Agents — intelligent workflows, automation, and real AI applications — built on AMD AI Developer Cloud, ROCm, and cloud-accessible AMD Instinct MI300X GPUs. Everything runs in the cloud; the focus is building, experimenting, benchmarking, and scaling AI workloads. Full challenge details drop at the kickoff.
Members of the AMD AI Developer Program get $100 in Dev Cloud credits, MI300X access, AMD AI Academy resources, a 1-month DeepLearning.AI Pro membership, and expert office hours.
We are building a peer-to-peer network of specialized Groot 1.7 expert nodes fused by a central Frankenstein Model Node. Each expert runs a large VLA model (e.g., fine-tuned LLaVA-based policies or OpenVLA) on an AMD Instinct MI300X instance (192 GB VRAM, ROCm 7.x). By utilizing each member's $100 AMD credit, we spin up a distributed cluster of specialized neural experts:
- Node A (Visual Search Expert): Camera observation analysis, frame decoding, and 3D coordinate detection on MI300X.
- Node B (Locomotion Expert): High-frequency joint angles and balance metrics for walking.
- Node C (Manipulation Expert): Bimanual hand and arm trajectory control for dexterous interaction.
At the center, the Frankenstein Model Node stitches these distributed experts into one brain: for each task goal + observation it fans out to the relevant experts and fuses their returned action logits into a single unified command (MoE-style late fusion / weighted gating). Every member's node connects to this Frankenstein node, creating a collaborative, decentralized robotic brain powered entirely by high-performance AMD hardware — no physics simulator in the loop.
Key components:
- Distributed VLA Inference — Multiple decentralized Groot 1.7 expert nodes served with PyTorch on ROCm.
- Frankenstein Model Node — Central fusion hub that stitches the distributed experts into one brain, routing each task to the relevant experts and fusing their action logits (MoE gating / weighted late fusion) into a unified command.
- Skill Co-Training & Adaptation — Saving successful trajectory data to dynamically fine-tune expert models on their respective MI300X GPUs via LoRA.
- Performance Harness — Throughput and latency benchmarking comparing FP16 vs INT8 execution on MI300X.
See ARCHITECTURE.md for the full system diagram.
Natural-language task goal + observation (image/state)
│
▼
┌────────────────────────┐
│ Frankenstein Model │
│ Node (Fusion Hub) │
│ route → fuse logits │
└───────────┬────────────┘
│ fan-out (REST/gRPC) ▲ fuse action logits
┌────────────────┼────────────────┬────────────┴───┐
▼ (Query Expert) ▼ (Query Expert) ▼ (Query Expert) ▼ (Query Expert)
┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐
│ Node A │ │ Node B │ │ Node C │ │ Node D │
│ (Visual │ │(Locomotion│ │ (Dexterous│ │ (Anomaly │
│ Search) │ │ Expert) │ │ Manipulation) │ Detection)│
└───────────┘ └───────────┘ └───────────┘ └───────────┘
────────────────────────────────────────────────────────────────
AMD Instinct MI300X GPU Fleet
(192 GB VRAM per Node · ROCm)
Full compiled brief — stack & access, prizes, submission fields, judging criteria — in docs/HACKATHON_REQUIREMENTS.md.
- Basic info: title · short + long description · technology & category tags
- Cover image · video presentation · slide presentation
- Public GitHub repository (MIT) · demo platform · application URL
- Signed up for the AMD AI Developer Program (unlocks credits + approval)
- Submitted before the deadline on the Schedule tab (Jul 11, 2026, 7:00 PM TRT at capture)
🚧 Enrolled, pre-kickoff. Concept: Decentralized VLA expert nodes fused by a central Frankenstein Model Node on AMD MI300X (ROCm) — no physics simulator. See docs/HACKATHON_REQUIREMENTS.md.
— Part of the Cadence multi-hackathon initiative.