Accelerating multi-contact modeling using a GPU

Multi-legged locomotion is classically modelled without accounting for contact slippage. Our group has previously demonstrated that such gaits are prone to sizeable slippage, showing evidence from both multi-legged organisms and robots. We have also shown a numerical ansatz based on viscous friction that rapidly provides an approximate solution. Here we report on advances we made in GPU acceleration of this computation, with the goal of demonstrating brute-force trajectory planning for a hexapedal robot with slipping. We search to find the best combination of leg motor commands to achieve a desired body velocity at each time-step - a 6 dimensional search space. We present the use of asynchronous data streams, device-based functions, local memory access and GPU-native sorting in NVIDIA CUDA using the Python Numba framework. Overall, we hope to expand this approach to enable real-time trajectory planning in multi-legged robots.