Bonus question: could this lead to a foundational model for Humanoid Control?
PULSE can randomly generate motion from noise and trained to perform different tasks using a sampler. So.....?
Now that PHC's code is out...
Introducing PULSE: Physics-based Universal motion Latent SpacE:
📜:
🌐:
All downstream tasks here use the same pretrained latent space. (1/6)
The key features of PULSE:
1. Once the latent space is learned, randomly sampled latents create stable and human-like behavior (instead of random jitters) -- better downstream exploration.
Here we visualize training for a "reach" and "move forward with speed" tasks. (2/6)
2. The latent space should have good **coverage**. In theory, it contains the latents for the motor skills required to perform all AMASS, so you are not crippled by using this latent space.
Here we use PULSE for free-from motion tracking (tracking VR controllers) (3/6)
3. The motor skills from the latent space should be able to extrapolate to unseen scenarios.
Here we show policy trained using PULSE can handle complex terrain traversal using human-like behavior, using only trajectory following reward (no additional adversarial reward) (4/6)
PULSE is distilled from a imitator (PHC+) that has achieved 100% success rate on AMASS.
It uses a variational information bottleneck (similar to a VAE) to form the latent space, jointly learning a prior that increases the expressiveness of the space. (5/6)