Der Physical-AI-Marktplatz

The PushT manipulation benchmark in LeRobot format — 206 teleop episodes.

The large-scale DROID Franka manipulation dataset in LeRobot v3 format — 95,658 episodes.

ALOHA bimanual simulation dataset — human-teleop cube transfer, 50 episodes.

ALOHA bimanual simulation dataset — human-teleop peg insertion, 50 episodes.

Cross-embodiment community pretraining dataset for VLA models — includes SO-100/SO-101 arms.

The LIBERO benchmark suites in LeRobot format — Panda simulation, 1,693 episodes.

SmolVLA base — the compact open vision-language-action policy for LeRobot.

Diffusion Policy trained on PushT — the classic LeRobot dataset/policy baseline pair.

ACT policy trained on the ALOHA cube-transfer dataset.

XVLA base vision-language-action policy in LeRobot format.

π0.5 base vision-language-action weights in LeRobot format (Gemma licence terms).

The official spec sheet for the Strix Halo APU — 16 Zen 5 cores, Radeon 8060S (40 CU, RDNA 3.5), XDNA 2 NPU (50+ TOPS), up to 128 GB LPDDR5X-8000 unified memory.

The community hub for running local LLMs on Strix Halo — containerised llama.cpp environments per backend, kept current with upstream.

Pre-built llama.cpp toolbox containers for gfx1151 — ROCm and Vulkan backends, rebuilt as llama.cpp master moves.

Toolbx-compatible containers for serving LLMs with vLLM on Strix Halo (gfx1151).

Open-source local LLM server with AMD optimisations — hybrid execution across the XDNA 2 NPU and Radeon GPU on Ryzen AI machines, OpenAI-compatible API.

Community wiki for local AI on Strix Halo — setup guides including llama.cpp with ROCm.