Festivus

policy

AMP (Adversarial Motion Priors)

UC Berkeley (Pieter Abbeel, Sergey Levine) · PyTorch

or hover any field below to flag it

Overview

Name
AMP (Adversarial Motion Priors)
Author
UC Berkeley (Pieter Abbeel, Sergey Levine)
Framework
PyTorch
License
bsd-3-clause
Skill type
locomotion
Evidence level
reported
Task description
Adversarial imitation learning for stylized locomotion. Motion capture clips train a discriminator providing style-reward signals while task rewards guide goal completion. Produces naturalistic gaits without manual reward shaping. Widely extended to real quadrupeds and humanoids via IsaacGym/IsaacLab.

Spaces

Action space
joint-position · 28-dim · 30Hz
Observation space
  • type: proprioception
  • · joint_angles
  • · joint_angular_velocities
  • · body_pos
  • · body_orientation
  • · body_linear_vel
  • · body_angular_vel

Links

HuggingFace repo
null
Paper (arXiv)
https://arxiv.org/abs/2104.02180

Compatible robots

1
G1Unitree Robotics

Compatible environments

2
flat-terrainnot in seedindoor-flatnot in seed

Datasets that reference this policy

0

No datasets reference AMP (Adversarial Motion Priors) yet.