LAFAN1

Ubisoft La Forge's public human motion-capture dataset, commonly used as a source of motion for humanoid robots to imitate.

LAFAN1 is an optical motion-capture dataset (optical motion capture: reflective markers on the body tracked by multiple cameras) released by Ubisoft's research division, La Forge, alongside its SIGGRAPH 2020 paper Robust Motion In-betweening. It contains 77 sequences from 5 actors, about 497,000 frames at 30 frames per second, totaling roughly 4.6 hours, with motions including walking, running, jumping, fighting, crawling, getting up after a fall, dancing, and climbing over obstacles; it's released as BVH skeletal animation files under the CC BY-NC-ND 4.0 license (attribution, non-commercial, no derivatives). It was originally meant for filling in in-between frames in game animation, but has been widely used in humanoid robotics in recent years: human motion is first mapped onto robot joints through motion retargeting, and a motion-tracking policy is then trained with reinforcement learning in simulation. Unitree has released a version of LAFAN1 retargeted onto its own humanoid robots, and the GMR retargeting tool directly supports its BVH files.

ExampleUC Berkeley's open-source BeyondMimic code trains its tracking policy directly on Unitree's retargeted LAFAN1 motions; the authors say any motion in the dataset suitable for a real robot can be trained directly with no parameter tuning.

Also called

Ubisoft La Forge Animation Dataset, LaFAN1

Related

AMASS (Archive of Motion Capture as Surface Shapes) · Optical Motion Capture · Motion Retargeting · General Motion Retargeting · BeyondMimic · Motion Tracking

Sources

Ubisoft La Forge Animation Dataset（LAFAN1）GitHub
GMR: General Motion Retargeting（GitHub）
BeyondMimic whole_body_tracking（GitHub）

As of

2020