Open-source projects that run on AMD ROCm. This page is only the map: code stays in each repo, longer write-ups stay in the tech notes and on rocpai-forge.github.io. The loop — capture, simulate, train, deploy — is drawn on the overview.
| Section | Open this if you want | |
|---|---|---|
| 1 | 🔥 TOP 3 | what just shipped |
| 2 | 💻 By machine | a short list for Strix Halo, MI300X, or a Radeon dGPU |
| 3 | 🎥 Real2Sim | scans, dynamic scenes, stereo depth |
| 4 | 🕹️ Simulation | teleoperation, datasets, MuJoCo |
| 5 | 🎯 RL and VLA | a policy that grasps, walks, or tracks velocity |
| 6 | 🤖 Robots | a desktop robot that talks, sees, or takes a photo |
| 7 | ⚡ Inference | a HIP engine you can compile |
| 8 | patches sent so other projects run on AMD GPUs | |
| 9 | 📚 2024–2025 | older guides, not retested in 2026-10 |
Three newest releases, newest first. A fourth one replaces the oldest. A LinkedIn post, when there is one, is linked on that item.
2026-10-07 · h3-hip.c v0.15.0. FastH3 paths are opt-in and off by default: --fasth3-lora (dense 4-step LoRA), --taeh3 (tiny video decoder), --vsa (sparse video attention). Prompt 1 at 1344×768, 5 seconds, one MI300X:
| Path | Time |
|---|---|
| dense 50-step | 668 s |
| FastH3 | 97 s |
| VSA + TAEH3 | 62 s |
FastH3 and VSA+TAEH3 are 4-step, not the published NVIDIA 50-step stack. At 62 s, one MI300X is faster than the published RTX 5090 full-opt cell (231 s) and the H100×4 baseline (81 s). Four H100s at full opt are still faster (23 s). gfx1151 and gfx90a were not re-timed for this tag. Ledger: MI300X vs NVIDIA, 768p. LinkedIn.
2026-08-29 · MicroDuck gait in UniLab. Pollen Robotics' Hugging Face mini biped. Velocity task on MuJoCo, PPO and SAC. Merged. LinkedIn: trained on a Radeon Pro W7900 with UniLab's native ROCm path.
2026-07-17 · SO-101 SimStudio. Teleop with a keyboard, Joy-Con, or leader arm, recorded as LeRobot v3.0. Collection release v0.1.2. Ubuntu 24.04 + ROCm, for a Ryzen AI laptop or mini PC. LinkedIn.
| Machine | Start here |
|---|---|
| Ryzen AI / Strix Halo (gfx1151, including Radeon 8060S) | FoundationStereo-hip.c · h3-hip.c · SO-101 SimStudio · ROCm Robotics RL Lab · Reachy Mini Chat |
| Instinct MI300X | h3-hip.c · OpenArm trajectories · OpenArm grasp RL · SO-101 Lab 01 · 4DGS · MicroDuck |
| Radeon dGPU | 4DGS viewer: Vulkan render, VA-API encode, WebRTC |
Scans, video, and stereo, turned into something a simulator can use.
| Project | What | Verified on |
|---|---|---|
| Scan2Sim | Scanned OBJ → MuJoCo asset with collision mesh, mass, and inertia. Repo runs. | Mesh pipeline. No GPU training. |
| Monocular video → 4DGS | One video → extra viewpoints → a dynamic Gaussian scene. Phi Media Lab × rocPAI-Lab. Note. 121 frames → 24 viewpoint streams → 30,000 updates. | Build on MI300X + ROCm. Viewer on Radeon: Vulkan + VA-API + WebRTC, 1280×720. |
| FoundationStereo-hip.c | HIP port of FoundationStereo for Strix Halo. Released v0.1.0. | gfx1151, ROCm 7.2. 2448×2048, 12 iterations: about 426 s in PyTorch ROCm, about 5.0 s in HIP. |
Teleoperation and the data those runs produce.
| Project | What | Verified on |
|---|---|---|
| SO-101 SimStudio | Keyboard, Joy-Con, or leader-arm teleop into LeRobot v3.0. Intro · deep dive. Collection release v0.1.2. Install path is Ubuntu 24.04 + ROCm. | Ubuntu 24.04 + ROCm, written for a Ryzen AI laptop or mini PC. |
| SO-101 Lab 01 | Same demonstrations, then ACT / SmolVLA, then closed-loop eval in MuJoCo. Note · deep dive. v0.1.3. | 50K-step checkpoint trained on MI300X. |
| OpenArm trajectories | Motion-planned pick-and-place demos for later VLA training. Grasp is either a calibrated top-down pose or a 6-DoF GraspGenX pose from ROCm. Note · deep dive. Data engine has been run. Replay is on CPU. | MI300X, ROCm 7.2, torch 2.7.1+rocm7.2 in the container. |
| RoboSimLab | SO-101 MuJoCo scene that exports LeRobot 3.0 episodes. Base scene and scripted export are done. IK, pick-and-place, and the real robot are still on the roadmap. | Target platform: Linux + PyTorch + ROCm. |
| OpenArm Labs Hub | Repo map and pitfalls for OpenArm sim, RL, imitation learning, and ROS. Code stays in the other repos. | — |
VLA training starts from SO-101 Lab 01 (ACT / SmolVLA) and the OpenArm trajectories above.
| Project | What | Verified on |
|---|---|---|
| OpenArm grasp RL | Cube grasp. The policy grew a grip the reward never spelled out. MuJoCo on CPU, policy learning on ROCm, training stack UniLab. Note · deep dive. rocPAI-Lab: Alex He, David Li, Andy Luo. | MI300X / MI210 + ROCm. |
| MicroDuck | PPO lab for an omnidirectional robot: reward, eval video, then keyboard control. Tasks live in microduck_rl_unilab. Chinese and English notebooks. | 500×300 timed at about 4 minutes on MI300X. |
| Kudada | 10-DoF biped velocity tracking on flat ground. FastSAC. Four accepted gaits: walk, crouch-walk, lateral walk, march-recover. Eval videos are in the repo. | UniLab + MuJoCo. |
| ROCm Robotics RL Lab | Panda lift-cube in robosuite, SAC / PPO. Sim rendering and PyTorch both use the iGPU. Train, eval, and sweep scripts are usable. | Ryzen AI Max+ 395, Radeon 8060S, PyTorch ROCm 7.1. |
| Project | What | Verified on |
|---|---|---|
| Reachy Mini Chat | Sim or real Reachy Mini: speech recognition, Ollama, TTS, emotion motion, lip sync. Can run fully offline. emo_v1–emo_v8, including Piper TTS. |
Ryzen AI Max+ 395 + Ubuntu 24.04. Also run on Ubuntu with a Radeon GPU. |
| ReachyBuddy | Same robot, more modes: voice photo, patrol with an Ollama vision model, chat, and an agent with tool calls. | — |
| ReachyClaw | Conversation and emotion motion through OpenClaw. The language model uses the OpenClaw API. The robot side is MuJoCo or the real robot. | — |
| VLM Demo | Browser camera answered by Qwen2.5-VL-3B-Instruct. Web demo, 2025-11. | A ROCm GPU. |
| Project | What | Verified on |
|---|---|---|
| h3-hip.c | HIP port of MiniMax-H3. One tree; set HIP_ARCH for the GPU. Actively updated. v0.15.0: on one MI300X, a 1344×768, 5-second video is about 668 s dense 50-step, about 62 s with VSA+TAEH3. |
gfx1151 (Strix Halo), gfx90a (MI210), gfx942 (MI300X). |
| dsh-plugin-h3-hip | Connects h3 --serve to DeepSeek Harness. Needs h3-hip.c ≥ v0.12-exp. |
— |
Pull requests sent to other projects so they can run on AMD GPUs. The list, with each pull request's GitHub state, is on the upstream page.
History keeps the older reproduction steps: LoRA / QLoRA on the W7900, Ollama on iGPU 780M, vLLM containers, EchoMimic, CosyVoice, the voice assistant, and RAG. Each entry records the GPU and ROCm version from that time. Not retested in 2026-10.
@misc{Playing with ROCm,
author = {He Ye (Alex)},
title = {Playing with ROCm},
howpublished = {\url{https://github.com/alexhegit/Playing-with-ROCm}},
year = {2024--2026}
}