Agile perceptive multiskill locomotion for quadrupedal robots in the wild

[00:00:02]We present a learning-based controller that enables quadrupedal robots to dynamically adapt their gates and skills and navigate both natural and urban environments like campuses and forests at high speeds using only on-board perception and computation.

[00:00:19]Quadrupedal animals move swiftly and naturally adjust their gait to handle different types of terrain. They trot for a balance of speed and endurance.

[00:00:29]When encountering a hurdle, they adaptively select a different gait to clear obstacle.

[00:00:34]Similar to animals, our robot uses a trotting gait when climbing down the stairs.

[00:00:39]In contrast, the robot chooses the bounding gait while jumping over hurdles.

[00:00:44]It also transitions its gates and skills while navigating obstacles. Moreover, our controller enables high-speed locomotion by quickly understanding the scene and responding accordingly. For instance, in this situation, the robot has only 0.68 seconds to react before jumping.

[00:01:02]We introduce action pre-trained transformer-based reinforcement learning, which consists of three stages: representation learning, reinforcement learning, and exteroceptive distillation.

[00:01:14]We first generate 15.5 hours of trotting and bounding motion torque data pair using trajectory optimization within 8 minutes.

[00:01:22]We then use a transformer variational autoencoder to extract essential state information from trotting and bounding motion data sets, representing them in a unified latent space.

[00:01:32]This learned latent space is then mapped to corresponding torque action from the data set through torque decoder.

[00:01:39]This decoder operates as an action prior, which we denoted as action pre-trained transformer. We reuse the learned action decoder to adapt to unseen 3D terrains for reinforcement learning in parallel robot environments.

[00:01:52]The actor outputs three components: a latent action, which is fed into the pre-trained decoder, an auxiliary action, which refines the decoded torque, and a gate selection logit, which chooses the appropriate gate decoder based on the current observation. Thanks to the decoder, which is already trained with physically plausible trajectories from a unified latent space, the robot exhibits flexible transitions and autonomously and adaptively selects between different gates and skills utilizing its observations. Also, we directly map the raw perception sensory data to policy input to reduce the gap between simulation and real world using teacher-student distillation.

[00:02:31]This enables our robot to directly use raw perception sensory data with minimal preprocessing in real world.

[00:02:38]The robot autonomously and adaptively selects gates and skills by utilizing its observations.

[00:02:44]During stair climbing, the robot selects a trot while ascending and simply walks when descending.

[00:02:50]However, after the stairs, it leaps down with a bounding gait.

[00:02:54]The robot automatically selects different gates depending on the step height, even when the commanded velocity is fixed.

[00:03:01]Similarly, on the same stairs, the robot utilizes a trot at low speeds but switches to bounding at higher speeds.

[00:03:09]On stepping stones, the robot exhibits similar behaviors.

[00:03:15]Although the decoder was trained solely on a flat 2D motion data set, its combination with the auxiliary architecture enabled the generation of diverse motions such as jumping, turning, and demonstrated robustness even under leg fracture conditions. For instance, before jumping over logs on flat terrain, since the motion is similar to 2D flat running, the torque from latent action remains dominant and operates in a consistent cycle.

[00:03:40]However, once the robot starts to jump, the auxiliary torque increases compared to the moment before the jump starts.

[00:03:50]Leveraging these capabilities in indoor experiments, the robot demonstrated agile locomotion such as transitioning gates at speeds exceeding 4 m/s and jumping over hurdles.

[00:04:02]Additionally, our robot reached a speed of 4.25 m/s while successfully jumping over 60 cm high steps, demonstrating stable and perceptive control performance.

[00:04:14]For outdoor deployment, the robot successfully navigated various structured and unstructured terrains, including grassy fields, rocks, uneven grounds, and extreme three-step staircases, reaching speeds of up to 6 m/s.

[00:04:40]We deployed our policy over wild forest trails, adapting speed and skills dynamically, testing various obstacle scenarios, such as vegetation, slopes, broken tree branch, gaps, logs at high speed.

[00:05:00]Thank you for watching.