AMASS (Archive of Motion Capture as Surface Shapes)

A large human-motion library that unifies 15 optical motion-capture datasets into the SMPL body-model format.

AMASS was proposed by Mahmood, Black, and colleagues at the Max Planck Institute for Intelligent Systems in Germany, published at ICCV 2019. Before it, different optical mocap datasets each used their own marker layouts and skeleton definitions, making them hard to combine. The authors used a method called MoSh++ to fit raw marker data to sequences of the SMPL parametric body model (which describes a person's pose and shape with a few dozen parameters), unifying 15 datasets into more than 300 subjects, over 11,000 motion sequences, and more than 40 hours in total. In embodied AI it is a standard source of human motion for training humanoid robots' motion tracking and whole-body control: SMPL motions are first retargeted into robot joint trajectories, then a policy is trained in simulation with reinforcement learning to track them. It requires registration and is intended for research use.

ExampleH2O retargeted about 13,000 motions from AMASS onto the Unitree H1, yielding 10,000 retargeted sequences, then used an imitation policy in simulation to filter out motions the robot physically couldn't perform, leaving about 8,500 sequences for training a whole-body humanoid teleoperation policy.

Also called

AMASS

Related

SMPL · Motion Retargeting · Motion Tracking · Optical Motion Capture · LAFAN1 · H2O

Sources

AMASS 官网 (Chinese)
Mahmood et al. 2019: AMASS (arXiv 1904.03278)
He et al. 2024: H2O: Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation

As of

2019-10