Learning Risk-Aware Quadrupedal Locomotion using Distributional Reinforcement Learning

Robotic Systems Lab are thrilled to unveil the latest breakthrough in the realm of robotics: Learning Risk-Aware Locomotion, as presented at ICRA 2024.

In this project, They have introduced a novel technique called Distributional Proximal Policy Optimization (DPPO), which empowers ANYmal robots to achieve locomotion with a keen awareness of risk factors.

Abstract: Traditional locomotion policies often overlook the inherent risks associated with robot movements. To address this gap, they propose a novel approach leveraging Distributional Proximal Policy Optimization (DPPO) to imbue robots, such as ANYmal quadrupeds, with risk-sensitive locomotion capabilities. This method allows the robot to navigate its environment while considering potential hazards, thereby enhancing safety and robustness in various scenarios.

Contributions:
- Introduction of Distributional Proximal Policy Optimization (DPPO) for risk-aware locomotion.
- Integration of DPPO into IsaacGym, enabling easy training of locomotion policies using various algorithms.
- Comprehensive comparison of locomotion algorithms including Distributed Distributional DDPG (D4PG), Deep Deterministic Policy Gradient (DDPG), Distributional PPO (DPPO), Distributional Soft Actor Critic (DSAC), Proximal Policy Optimization (PPO), Soft Actor Critic (SAC), and Twin Delayed DDPG (TD3).

Resources:
- Video Presentation: Watch Here - https://youtu.be/GGFXpF4qeVY?si=z_WgfF3LGPdA6SU9
- Arxiv Paper: Read Here - https://arxiv.org/abs/2309.14246
- GitHub Repository: Access Code Here - https://github.com/leggedrobotics/rsl_rl/tree/algorithms
- Project Page: Explore Here - https://sites.google.com/leggedrobotics.com/risk-aware-locomotion/home
- Authors: Lukas Schneider, Jonas Frey, Takahiro Miki, Marco Hutter

Robotic Systems Lab invite researchers and enthusiasts to delve into the work, explore the codebase, and contribute to the advancement of risk-aware locomotion in robotics. Together, pave the way for safer and more efficient robot mobility in diverse environments.