Core Concepts
FootstepNet provides efficient footstep planning and forecasting for bipedal robots using Deep Reinforcement Learning techniques.
Abstract
I. Introduction
Designing humanoid locomotion controllers is challenging.
Footstep planning is crucial for navigating in local environments.
Existing methods rely on search-based algorithms or hand-crafted tuning.
II. Problem Statement
Footstep planning aims to find a suitable sequence of footsteps efficiently and safely.
Trajectories are computed based on simplified models like 3D-LIPM.
III. Background on RL and DRL
Reinforcement Learning (RL) involves learning policies to maximize rewards.
Deep RL algorithms use neural networks to approximate policies and value functions.
IV. Method
Defines footstep parameters, state-space, action-space, reward function, and termination criteria.
Utilizes an MDP formulation for footstep planning with continuous actions.
V. Experiments
Compares FootstepNet planner with ARA* planner in various scenarios.
Validates the accuracy of FootstepNet forecast for predicting the number of steps required to reach a target.
VI. Conclusion
FootstepNet outperforms ARA* planner in terms of efficiency and performance.
Demonstrates effectiveness in real-world scenarios like RoboCup 2023 competition.
Stats
"The execution time of FootstepNet planner is 45µs per footstep."
"In the worst case, the RL agent is equal or better in 97.2% of the experiments."