HIM

A legged-locomotion method that uses only proprioception, inferring terrain and disturbances from the robot's own response.

HIM was released in December 2023 by Shanghai AI Lab's OpenRobotLab (Jiangmiao Pang's group), published at ICLR 2024, with its code repository named HIMLoco. A legged robot's sensors only ever give incomplete, noisy observations, and external state such as ground friction and terrain height is hard to estimate directly. HIM borrows the idea of internal model control from classical control theory, treating this external state as a disturbance to be inferred from the robot's own response: a “hybrid internal embedding” represents both explicit body velocity and implicit stability information at once, and contrastive learning is used to pull it close to the robot's actual next-moment state. It needs only proprioception from joint encoders and the IMU, skipping the usual two-stage teacher-student imitation process; the paper reports that training for about an hour on a single RTX 4090 is enough to let a quadruped handle a variety of terrains and external-force disturbances.

ExampleA quadruped relying only on joint encoders and the IMU can keep walking under terrain and external shoves it never saw during training.

Also called

Hybrid Internal Model, HIMLoco, Learning Agile Legged Locomotion with Simulated Robot Response

Related

Legged Locomotion · Proprioception · Blind Locomotion · Contrastive Learning · Rapid Motor Adaptation · Shanghai Artificial Intelligence Laboratory

Sources

Hybrid Internal Model (arXiv 2312.11460)
OpenRobotLab/HIMLoco (GitHub)

As of

2024-01