Quadrupedal Landing Control Under Low Gravity

Quadrupedal attitude control & landing control under reduced gravity with diverse initial poses, trained using Proximal Policy Optimization (PPO)


Ubuntu Mujoco PyTorch NumPy JAX wandb


Link to Github Repo

Demo

  • Attitude Control

    reorientation
  • Landing Control

    landing

Description

  • Mitigated the computational intractability of fine-grained particle simulation of granular terrains by implementing the vertical resistive force component of a granular media contact model within MuJoCo, enabling rigid-body approximation of loose-sand impact dynamics
  • Leveraged MuJoCo MJX (JAX) and MuJoCo Playground for mass environment vectorization on GPU, enabling high-throughput Proximal Policy Optimization (PPO) training that generalizes across arbitrary task configurations via domain randomization
  • Trained a zero-gravity attitude control policy for Unitree Go1 quadruped to execute precise aerial reorientation that exploits the conservation of angular momentum to dynamically redistribute rotational inertia
  • Demonstrated robust landing capabilities by integrating attitude control with impact recovery; the learned policy leverages articulated compliance to dissipate impact energy, achieving stable four-foot touchdown on the granular terrain from arbitrary initial orientations