Robot Perception and Control

Learning by Cheating arxiv github

Proposed two-stage training procedure, which first train a privileged agent and then using the agent as a teacher to train a purely vision-based system, for effective imitation learning. This paradigm is the underlying concept in the legged RL.

Learning to Walk in Minutes arvix github

Presents a training setup that achieves fast policy generation for real-world robotic tasks by using massive parallelism on a single workstation GPU (showcase of Isaac Gym).

• A codebase is widely used as baseline for developing legged locomotion system.

Walk These Ways arxiv

Perceptive locomotion for quadrupeds

Presented a three stage training and deploy method to perform zero-shot sim-to-real transfer [1].

1. a teacher policy, which has access to privileged information, is trained to follow a random target velocity over randomly generated terrain with random disturbances.
2. a student policy is trained to reproduce the teacher policy’s actions without using this privileged information.
3. transfer the learned student policy to the physical robot and deploy it in the real world with onboard sensors.

Training teacher policy

Training student policy

Deployment

Legged Locomotion using Egocentric Vision arxiv

Parkour Learning arxiv

Humanoid Parkour Learning arxiv