Page 99 - Read Online

Page 290 Zhang et al. Intell Robot 2022;2(3):275­97 I http://dx.doi.org/10.20517/ir.2022.20

ANYmal Minitaur Spot Mini ANYmal Unitree Laikago Unitree Laikago Daisy Hexapod Unitree- Laikago

and
RaiSim Pybullet Pybullet RaiSim, PyBullet MuJoCo. IsaacGym [102] Pybullet PyBullet PyBullet

Velocities, Foot Speed, Recov- HAA, In- Impulse, Flu- and Yaw, (2-dim), (3-dim). Velocity, Acceleration and Velocity, Velocity. Tar- to Duration. Orienta- Refer- vs. State, and Base

Base Joint Fluency. Motion, Joint and Slip Direction, and Orientation Velocities Angular Foot Smoothness, Linear Base Linear (Base Distance Timestep and Distance (Current Velocities Positions, Velocity).

Locomotion: Torque, Cost, Direction, State, Torque, ery: KFE, HFE, Contact, ternal ency. Position Base Smoothness. Distance, Angular Base and Linear Torque, Slip, and Orientation. Torques, Base Desired Euclidean and get) Forward Deviation. tion Gap Motion Joint in ence End-Effector and Pose

Joint Joint Residu- Joint Primitive Vector Frequency, Stride Residual Joint De- for

Desired Positions. Desired Positions. Position Foot als. Desired Positions. One-Hot Selection (9-dim) Gait Height, Swing Length, Action. Desired Positions. Torques Joint Positions. sired

Joint IMU Previous Angu- Linear Foot Ori- Out- positions Primitive Orientation, Tar- Pre- Genera- and Se-

Velocities, State(withHistory),PreviousAc- Gravity. (6-step), and Base (2-dim), and (3-dim) and (3-dim), Velocities, Policy State, Velocity. linear Env., Angles, Joint Direction, Trajectory (3-dim) Action and

Base Command, Angles (6-step), (6-step). Velocities and Height Joint Base Desired (without positions), (previously-used). to Velocity, and Distance Action, (1-dim) (Orientation Rotations)

Height, tion, Motor Readings Action Orientation lar Accelerations Phase. Base entation, put, Pose foot and Distance Base get vious Param. tor Phase Pose Joint quence.

TRPO SAC ARS [23] GCPO DQN PPO SAC PPO

lever- and generation learning locomotion the in human frame- dy- with randomized ter- framework commands com- Control robust ter- for ground-truth generaliz- for on skills robot skills animals.

method automated data for RL Deep minimal Sim2Real domain uneven RL velocity framework Model-Based synthesize controllers. approach rain locomotion using exterocep- without and framework locomotion legged locomotion real-world

Sim2Real fast, cost-effective system quadrupedal with with world quadrupedal work utilizing offline RL and traversing allow CPPO-based tracking constraints. hierarchical to RL quadrupedal learning inputs maps. sample-efficient hierarchical robots. real-world enabling system agile learn imitating

A aging schemes. A policies real effort. A namics to rain. A for under A bining and A tive height A able learning A to by

2019 2020 2020 2020 2020 2020 2020
Science Robotics CoRL ArXiv IEEE Robotics Autom CoRL ArXiv ICRA 2020 RSS

motor World Random- Bezier Legged Optimiza- Robot Con- Lo- Legged for Locomotion Reinforce- Locomotion

dynamic Real the Effort [41] Domain with Policy Quadrupedal Contact-Adaptive Efficient Generalization Policies [103] Animals [10]

and robots [6] in Human Modulation Sim-to-Real Generalizable Hierarchical Robotic
agile legged Walk to Minimal and for Constrained Dynamic a Robust, Terrain Locomotion with Learning [47] Agile Imitating

Learning for skills Learning with Dynamics Gait ized Curves Locomotion [50] Guided for tion Locomotion [28] Learning for troller comotion [101] Zero-Shot Visual Learning Skills ment Learning by Skills