HumanoidMimicGen Generates Loco-Manipulation Data for Humanoid Robots
A team of researchers has introduced HumanoidMimicGen, a novel technique for the automatic creation of data related to humanoid legged loco-manipulation. Collecting the numerous demonstrations needed for imitation learning through teleoperation is a labor-intensive process. While current algorithms are effective for manipulators, they struggle with humanoids due to the complexities of high-dimensional action spaces involving arms, legs, and torsos. HumanoidMimicGen enhances whole-body skills that are rich in contact by adapting a limited number of source demonstrations to various new states, effectively generalizing with changes in object poses. By integrating single- and dual-arm skills with whole-body locomotion and manipulation, it produces stable, collision-free data suitable for various environments. Additionally, a new simulated benchmark for loco-manipulation has been established to assess this method.
Key facts
- HumanoidMimicGen is a method for generating humanoid legged loco-manipulation data.
- It adapts contact-rich whole-body skills from a handful of source demonstrations.
- The method generalizes across changes in object pose.
- It interleaves single- and dual-arm skills with whole-body locomotion and manipulation planning.
- The generated data is stable and collision-free across diverse scenes and layouts.
- A new simulated loco-manipulation benchmark was introduced for evaluation.
- Existing data-generation algorithms are ineffective on humanoids due to high-dimensional composite action spaces.
- Imitation learning requires a large number of demonstrations, which are difficult to collect via teleoperation.
Entities
—