ASE

Berkeley and NVIDIA's 2022 method for learning a reusable latent space of skills from motion-capture clips.

ASE was proposed by Xue Bin Peng, Sergey Levine, Sanja Fidler, and colleagues at UC Berkeley and NVIDIA, published at SIGGRAPH 2022. In simulated character animation, every new task usually means training a policy from scratch, relearning basic motions like walking and running again and again. ASE works in two layers. During pretraining, on a large batch of unlabeled, unsegmented motion clips, it trains a low-level policy conditioned on a latent variable z, using adversarial imitation learning (a discriminator judges whether a motion looks human, as in the data) to keep motions natural, and unsupervised skill discovery so that different values of z correspond to different skills. For a downstream task, the low-level policy is frozen and only a high-level policy that outputs z is trained, which needs only a simple reward. Using parallel simulation in Isaac Gym, the low-level policy was trained on more than 10 billion samples. It's a follow-up to AMP (Adversarial Motion Priors).

ExampleA simulated humanoid character carrying a sword and shield first pretrains its skill latent space on about 30 minutes of motion data across 187 clips; afterward, given only a reward for “run to the target and knock it down,” the high-level policy can compose running, sword swings, and a strike into a coherent sequence.

Also called

Adversarial Skill Embeddings, ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters

Related

Adversarial Motion Priors · Generative Adversarial Imitation Learning · Unsupervised Skill Discovery · Latent Space · Behavior Foundation Model · Isaac Gym

Sources

ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters (arXiv 2205.01906)
ASE 项目页（Xue Bin Peng） (Chinese)

As of

2022-07