HoST (Humanoid Standing-up)

A control framework that uses reinforcement learning to learn, from scratch, how a humanoid can stand up from many different postures.

HoST was released in February 2025 by Shanghai AI Lab (Jiangmiao Pang's group) together with Shanghai Jiao Tong University, the University of Hong Kong, Zhejiang University, and CUHK, published at RSS 2025 and nominated for best systems paper. It addresses how a humanoid robot gets itself back up after falling: earlier methods either ran only in simulation and ignored real motor limits, or relied on standing-up trajectories hand-designed for a specific ground surface. HoST learns from scratch with reinforcement learning across a variety of simulated terrains, using no reference motion at all; it uses multiple value networks (critics), each evaluating a different group of rewards, paired with curriculum learning, plus smoothness regularization and an implicit velocity cap to prevent jitter and overly forceful motion on the real robot. Once trained, it was deployed directly to a Unitree G1, able to stand up from postures like lying down or leaning against a wall in a variety of indoor and outdoor settings, and is a representative work in fall-recovery research.

ExampleWhether a Unitree G1 is lying on the ground or leaning against a wall, the HoST policy can control it to stand up on its own, with no human support or overhead rig needed.

Also called

Learning Humanoid Standing-up Control across Diverse Postures

Related

Fall Recovery · RL-based Locomotion Control · Curriculum Learning · Sim-to-Real Transfer · Unitree G1 · Shanghai Artificial Intelligence Laboratory

Sources

Learning Humanoid Standing-up Control across Diverse Postures (arXiv 2502.08378)
HoST project page

As of

2025-04