OLIVE: Online Low-Rank Incremental Learning for Efficient Adaptive Exoskeletons
Paper • 2606.05234 • Published
Frozen base policy for OLIVE: a compact multimodal controller mapping on-body sensors (IMU, joints, EMG, vibration, context) to bilateral hip torques. Distilled from Physical Intelligence π₀.₅ / openpi and kept frozen during online low-rank adaptation.
Paper: https://huggingface.co/papers/2606.05234
Dong Liu, Yanxuan Yu, Ben Lengerich, Tong Geng, Ying Nian Wu
| File | Description |
|---|---|
pytorch_model.bin |
PyTorch BaseController state dict |
base_controller_w0.bin |
C++ runtime layout (OLIVEModel::load_base_weights) |
config.json |
Architecture hyperparameters |
import torch
from distillation.student import BaseController
model = BaseController()
model.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"))
model.eval()
./olive_deploy base_controller_w0.bin