Navigating slippery terrain via domain randomization

Student: Florian Tanguy
Supervision: Tingting Ni, Kai Ren

Navigation performance on slippery terrain using different policies:
1. Memoryless Feedforward MLP
2. Three-step frame-stacking MLP
3. Long short-term memory

Develop a data-driven domain randomization framework to enable a wheeled robot to safely navigate in a maze under wet wheels and slippery terrain. A Gaussian Mixture Model over the wheel gain and deadzone is identified from data to model multi-modal dynamics (“slippery” vs. “grippy” regimes). A Recurrent Neural Network (LSTM) policy outperformed a memoryless feedforward network and frame stacking with MLP.