Kraftwerk Bot: trained models
Trained weights and results for github.com/diw22/DWinters_Humanoid: an
audio-conditioned OpenVLA-OFT that hears a drum loop with one part missing and plays that part on a simulated Unitree
G1. Download into the code's runs/ folder:
hf download diw22/DWinters_Humanoid-models --local-dir runs
| Folder | Model |
|---|---|
three_part/ |
OpenVLA-OFT + MERT audio, trained on claps, rims and shakers (LoRA adapters, audio and state projectors, action head) |
rims_only/ |
three_part fine-tuned on rims alone |
rims_student/ |
a 1.8M-parameter student distilled from rims_only (runs on a CPU) |
jaco_v1/ |
results of version 1 (claps on two simulated Jaco arms) |
Each folder also has config.json (normalisation and training settings), metrics.jsonl (training curve) and
evaluation_test.json (closed-loop results on the held-out takes, MuJoCo 3.3.0); the 7B models have test videos.
Results and how to reproduce them are in the GitHub README.
The 7B models are LoRA adapters and heads for openvla/openvla-7b, which builds on Llama 2: use them under OpenVLA's and Llama 2's terms. The audio encoder, MERT-v1-95M, is not included; the code downloads it.
Model tree for diw22/DWinters_Humanoid-models
Base model
openvla/openvla-7b