How mbrossar/rins-w works
Project type and purpose
• This is an open-source scientific/robotics tool (not a web app or full-stack system). It implements RINS-W: a real-time inertial navigation system on wheeled robots. It has two main functions:
• Offline training of a deep neural network to detect zero‐velocity events from IMU data
• Online/inference fusion of IMU measurements and zero‐velocity detections via an iterated extended Kalman filter (IEKF)
File structure
• Top‐level directories and files:
• data/Kaist – raw IMU datasets (pickled)
• results/Kaist – trained model checkpoints, hyperparameters, weights
• src/ – library modules: dataset.py, iekf.py, learning.py, lie_algebra.py, losses.py, networks.py, utils.py
• main_Kaist.py & main_Kalman_Kaist.py – entrypoints for training & testing/inference
• dataset.py, iekf.py, etc. mirrored at root—likely quick scripts or legacy
• requirements.txt – dependency list (PyTorch, numpy, etc.)
README for additional insights
• Core components:
• “adapter block” = deep neural network (dilated convolutions + CNNs)
• Kalman filter that uses zero‐velocity detector outputs and pseudo‐measurements (no lateral slip, zero vertical vel.)
• Tech stack: Python 3.5+, PyTorch 1.5
Detailed component breakdown
• Data Layer (cyan):
• data/Kaist/*.p → src/dataset.py → training engine
• Training Engine (orange):
• src/learning.py integrates dataset, networks, losses → runs backprop → writes results/Kaist/YYYY…
• results/Kaist/.../weights.pt, net_params.yaml, hparams.yaml
• Inference Engine (purple):
• main_Kalman_Kaist.py loads IMU feed & model → src/iekf.py + src/utils.py → publishes state estimates
• Lie‐Algebra & Filter Core (gray):
• src/lie_algebra.py + src/iekf.py hold mathematical routines for propagation & update